# AI Runtime Security Tools: every page as markdown Sixteen AI runtime security tools scored on what they stop before it completes, where they enforce, latency, deployment and proof. Read 16 Nov 2026. --- # AI runtime security tools, ranked by the actions they stop URL: https://airuntimesecuritytools.com/ Updated: 2026-11-16 Sixteen AI runtime security tools are scored on 16 Nov 2026 from 896 facts in their evidence files. HiddenLayer leads the site ranking with 71.0 points. Bay is joint 4th with 62.5. Pick an action below to see which tools stop it before it completes. ## The short answer HiddenLayer leads the site ranking at 71.0 of 100. Bay is joint 4th at 62.5. HiddenLayer is ahead mainly on layers covered (9 against 6, worth 6.0 points at weight 20). ## Which tools stop this action Each row is an action an AI agent or app can take. The table lists the vendors that document stopping it before it completes, the ones that document it in part (for some agents only, or detect and alert without blocking), and the ones with nothing found on 16 Nov 2026. With scripts on, tick the actions you need stopped and the result below lists the tools documented to stop all of them. **Actions and which vendors stop them before they complete** | Action | Documented | Partial | None found | |---|---|---|---| | Shell command | [Bay](/vendors/bay/#fact-bay-48), [Cisco](/vendors/cisco/#fact-cisco-10), [HiddenLayer](/vendors/hiddenlayer/#fact-hiddenlayer-10), [Operant AI](/vendors/operant-ai/#fact-operant-ai-26), [Straiker](/vendors/straiker/#fact-straiker-08) | [CrowdStrike](/vendors/crowdstrike/#fact-crowdstrike-09), [Lasso Security](/vendors/lasso-security/#fact-lasso-security-13), [Microsoft](/vendors/microsoft/#fact-microsoft-19), [Pillar Security](/vendors/pillar-security/#fact-pillar-security-09), [Zenity](/vendors/zenity/#fact-zenity-50) | [Aim Security](/vendors/aim-security/#fact-aim-security-62), [Check Point](/vendors/check-point/#fact-check-point-44), [Cloudflare](/vendors/cloudflare/#fact-cloudflare-24), [Noma Security](/vendors/noma-security/#fact-noma-security-50), [Palo Alto Networks](/vendors/palo-alto-networks/#fact-palo-alto-networks-69), [SentinelOne (Prompt Security)](/vendors/sentinelone/#fact-sentinelone-53) | | File access | [HiddenLayer](/vendors/hiddenlayer/#fact-hiddenlayer-12), [Pillar Security](/vendors/pillar-security/#fact-pillar-security-10), [Straiker](/vendors/straiker/#fact-straiker-08) | [Bay](/vendors/bay/#fact-bay-24), [Check Point](/vendors/check-point/#fact-check-point-46), [Cisco](/vendors/cisco/#fact-cisco-20), [CrowdStrike](/vendors/crowdstrike/#fact-crowdstrike-09), [Lasso Security](/vendors/lasso-security/#fact-lasso-security-12), [Microsoft](/vendors/microsoft/#fact-microsoft-19), [Operant AI](/vendors/operant-ai/#fact-operant-ai-24), [Zenity](/vendors/zenity/#fact-zenity-28) | [Aim Security](/vendors/aim-security/#fact-aim-security-62), [Cloudflare](/vendors/cloudflare/#fact-cloudflare-24), [Noma Security](/vendors/noma-security/#fact-noma-security-50), [Palo Alto Networks](/vendors/palo-alto-networks/#fact-palo-alto-networks-69), [SentinelOne (Prompt Security)](/vendors/sentinelone/#fact-sentinelone-53) | | Data paste | [Aim Security](/vendors/aim-security/#fact-aim-security-23), [Bay](/vendors/bay/#fact-bay-23), [Check Point](/vendors/check-point/#fact-check-point-13), [CrowdStrike](/vendors/crowdstrike/#fact-crowdstrike-36), [Lasso Security](/vendors/lasso-security/#fact-lasso-security-17), [Microsoft](/vendors/microsoft/#fact-microsoft-10), [Operant AI](/vendors/operant-ai/#fact-operant-ai-28), [Palo Alto Networks](/vendors/palo-alto-networks/#fact-palo-alto-networks-20), [SentinelOne (Prompt Security)](/vendors/sentinelone/#fact-sentinelone-19) | [Cisco](/vendors/cisco/#fact-cisco-48), [Cloudflare](/vendors/cloudflare/#fact-cloudflare-08), [Noma Security](/vendors/noma-security/#fact-noma-security-20), [Pillar Security](/vendors/pillar-security/#fact-pillar-security-26), [Straiker](/vendors/straiker/#fact-straiker-10) | [HiddenLayer](/vendors/hiddenlayer/#fact-hiddenlayer-09), [Zenity](/vendors/zenity/#fact-zenity-39) | | Package install | [CrowdStrike](/vendors/crowdstrike/#fact-crowdstrike-16), [Operant AI](/vendors/operant-ai/#fact-operant-ai-27), [Zenity](/vendors/zenity/#fact-zenity-51) | none | [Aim Security](/vendors/aim-security/#fact-aim-security-62), [Bay](/vendors/bay/#fact-bay-57), [Check Point](/vendors/check-point/#fact-check-point-44), [Cisco](/vendors/cisco/#fact-cisco-52), [Cloudflare](/vendors/cloudflare/#fact-cloudflare-24), [HiddenLayer](/vendors/hiddenlayer/#fact-hiddenlayer-10), [Lasso Security](/vendors/lasso-security/#fact-lasso-security-12), [Microsoft](/vendors/microsoft/#fact-microsoft-19), [Noma Security](/vendors/noma-security/#fact-noma-security-50), [Palo Alto Networks](/vendors/palo-alto-networks/#fact-palo-alto-networks-69), [Pillar Security](/vendors/pillar-security/#fact-pillar-security-09), [SentinelOne (Prompt Security)](/vendors/sentinelone/#fact-sentinelone-53), [Straiker](/vendors/straiker/#fact-straiker-43) | | Network call | [Cisco](/vendors/cisco/#fact-cisco-14), [Straiker](/vendors/straiker/#fact-straiker-23), [Zenity](/vendors/zenity/#fact-zenity-48) | [CrowdStrike](/vendors/crowdstrike/#fact-crowdstrike-24), [HiddenLayer](/vendors/hiddenlayer/#fact-hiddenlayer-08), [Microsoft](/vendors/microsoft/#fact-microsoft-20), [Noma Security](/vendors/noma-security/#fact-noma-security-51), [Operant AI](/vendors/operant-ai/#fact-operant-ai-30), [Pillar Security](/vendors/pillar-security/#fact-pillar-security-10) | [Aim Security](/vendors/aim-security/#fact-aim-security-62), [Bay](/vendors/bay/#fact-bay-57), [Check Point](/vendors/check-point/#fact-check-point-44), [Cloudflare](/vendors/cloudflare/#fact-cloudflare-24), [Lasso Security](/vendors/lasso-security/#fact-lasso-security-12), [Palo Alto Networks](/vendors/palo-alto-networks/#fact-palo-alto-networks-69), [SentinelOne (Prompt Security)](/vendors/sentinelone/#fact-sentinelone-53) | | Credential access | [Bay](/vendors/bay/#fact-bay-48), [Operant AI](/vendors/operant-ai/#fact-operant-ai-24), [Pillar Security](/vendors/pillar-security/#fact-pillar-security-10), [Zenity](/vendors/zenity/#fact-zenity-36) | [Check Point](/vendors/check-point/#fact-check-point-28), [Cisco](/vendors/cisco/#fact-cisco-15), [HiddenLayer](/vendors/hiddenlayer/#fact-hiddenlayer-09), [Lasso Security](/vendors/lasso-security/#fact-lasso-security-23), [Microsoft](/vendors/microsoft/#fact-microsoft-30), [Noma Security](/vendors/noma-security/#fact-noma-security-52), [Straiker](/vendors/straiker/#fact-straiker-08) | [Aim Security](/vendors/aim-security/#fact-aim-security-62), [Cloudflare](/vendors/cloudflare/#fact-cloudflare-24), [CrowdStrike](/vendors/crowdstrike/#fact-crowdstrike-62), [Palo Alto Networks](/vendors/palo-alto-networks/#fact-palo-alto-networks-69), [SentinelOne (Prompt Security)](/vendors/sentinelone/#fact-sentinelone-53) | | Prompt injection | [Aim Security](/vendors/aim-security/#fact-aim-security-59), [Bay](/vendors/bay/#fact-bay-51), [Check Point](/vendors/check-point/#fact-check-point-42), [Cisco](/vendors/cisco/#fact-cisco-03), [Cloudflare](/vendors/cloudflare/#fact-cloudflare-01), [CrowdStrike](/vendors/crowdstrike/#fact-crowdstrike-13), [HiddenLayer](/vendors/hiddenlayer/#fact-hiddenlayer-07), [Lasso Security](/vendors/lasso-security/#fact-lasso-security-18), [Microsoft](/vendors/microsoft/#fact-microsoft-19), [Noma Security](/vendors/noma-security/#fact-noma-security-22), [Operant AI](/vendors/operant-ai/#fact-operant-ai-23), [Palo Alto Networks](/vendors/palo-alto-networks/#fact-palo-alto-networks-47), [Pillar Security](/vendors/pillar-security/#fact-pillar-security-25), [SentinelOne (Prompt Security)](/vendors/sentinelone/#fact-sentinelone-41), [Straiker](/vendors/straiker/#fact-straiker-22), [Zenity](/vendors/zenity/#fact-zenity-52) | none | none | | Harmful output | [Aim Security](/vendors/aim-security/#fact-aim-security-61), [Check Point](/vendors/check-point/#fact-check-point-42), [Cisco](/vendors/cisco/#fact-cisco-02), [Cloudflare](/vendors/cloudflare/#fact-cloudflare-06), [HiddenLayer](/vendors/hiddenlayer/#fact-hiddenlayer-09), [Lasso Security](/vendors/lasso-security/#fact-lasso-security-14), [Palo Alto Networks](/vendors/palo-alto-networks/#fact-palo-alto-networks-56), [SentinelOne (Prompt Security)](/vendors/sentinelone/#fact-sentinelone-41), [Straiker](/vendors/straiker/#fact-straiker-40) | [Bay](/vendors/bay/#fact-bay-52), [CrowdStrike](/vendors/crowdstrike/#fact-crowdstrike-18), [Microsoft](/vendors/microsoft/#fact-microsoft-43), [Noma Security](/vendors/noma-security/#fact-noma-security-22), [Operant AI](/vendors/operant-ai/#fact-operant-ai-22), [Pillar Security](/vendors/pillar-security/#fact-pillar-security-26), [Zenity](/vendors/zenity/#fact-zenity-53) | none | Tick one or more actions above. ## The ranking [The full ranking](/ranking/) has the same table with re-rank sliders. Each score links to the evidence on the vendor's page. **Site ranking: scores out of 10 per criterion, totals out of 100** | Rank | Vendor | Total (of 100) | Action types stopped before they complete (weight 30) | Layers covered (weight 20) | Latency and failure mode (weight 15) | Deployment (weight 15) | Detection and response (weight 5) | Transparency (weight 10) | Proof (weight 5) | |---|---|---|---|---|---|---|---|---|---| | 1st | [HiddenLayer](/vendors/hiddenlayer/) | 71.0 | [8](/vendors/hiddenlayer/#evidence-stops) | [9](/vendors/hiddenlayer/#evidence-layers) | [4](/vendors/hiddenlayer/#evidence-latency) | [6](/vendors/hiddenlayer/#evidence-deploy) | [9](/vendors/hiddenlayer/#evidence-response) | [6](/vendors/hiddenlayer/#evidence-transparency) | [7](/vendors/hiddenlayer/#evidence-proof) | | 2nd | [Cisco](/vendors/cisco/) | 66.5 | [7](/vendors/cisco/#evidence-stops) | [8](/vendors/cisco/#evidence-layers) | [5](/vendors/cisco/#evidence-latency) | [7](/vendors/cisco/#evidence-deploy) | [8](/vendors/cisco/#evidence-response) | [5](/vendors/cisco/#evidence-transparency) | [5](/vendors/cisco/#evidence-proof) | | 3rd | [Lasso Security](/vendors/lasso-security/) | 63.5 | [7](/vendors/lasso-security/#evidence-stops) | [7](/vendors/lasso-security/#evidence-layers) | [5](/vendors/lasso-security/#evidence-latency) | [7](/vendors/lasso-security/#evidence-deploy) | [7](/vendors/lasso-security/#evidence-response) | [3](/vendors/lasso-security/#evidence-transparency) | [8](/vendors/lasso-security/#evidence-proof) | | joint 4th | [Bay](/vendors/bay/) | 62.5 | [8](/vendors/bay/#evidence-stops) | [6](/vendors/bay/#evidence-layers) | [5](/vendors/bay/#evidence-latency) | [8](/vendors/bay/#evidence-deploy) | [7](/vendors/bay/#evidence-response) | [0](/vendors/bay/#evidence-transparency) | [7](/vendors/bay/#evidence-proof) | | joint 4th | [Operant AI](/vendors/operant-ai/) | 62.5 | [7](/vendors/operant-ai/#evidence-stops) | [7](/vendors/operant-ai/#evidence-layers) | [4](/vendors/operant-ai/#evidence-latency) | [7](/vendors/operant-ai/#evidence-deploy) | [7](/vendors/operant-ai/#evidence-response) | [4](/vendors/operant-ai/#evidence-transparency) | [7](/vendors/operant-ai/#evidence-proof) | | joint 4th | [Straiker](/vendors/straiker/) | 62.5 | [7](/vendors/straiker/#evidence-stops) | [7](/vendors/straiker/#evidence-layers) | [5](/vendors/straiker/#evidence-latency) | [7](/vendors/straiker/#evidence-deploy) | [5](/vendors/straiker/#evidence-response) | [3](/vendors/straiker/#evidence-transparency) | [8](/vendors/straiker/#evidence-proof) | | 7th | [Palo Alto Networks](/vendors/palo-alto-networks/) | 58.5 | [6](/vendors/palo-alto-networks/#evidence-stops) | [7](/vendors/palo-alto-networks/#evidence-layers) | [4](/vendors/palo-alto-networks/#evidence-latency) | [6](/vendors/palo-alto-networks/#evidence-deploy) | [4](/vendors/palo-alto-networks/#evidence-response) | [6](/vendors/palo-alto-networks/#evidence-transparency) | [7](/vendors/palo-alto-networks/#evidence-proof) | | 8th | [Pillar Security](/vendors/pillar-security/) | 55.5 | [7](/vendors/pillar-security/#evidence-stops) | [7](/vendors/pillar-security/#evidence-layers) | [0](/vendors/pillar-security/#evidence-latency) | [6](/vendors/pillar-security/#evidence-deploy) | [7](/vendors/pillar-security/#evidence-response) | [4](/vendors/pillar-security/#evidence-transparency) | [8](/vendors/pillar-security/#evidence-proof) | | 9th | [Check Point](/vendors/check-point/) | 55.0 | [5](/vendors/check-point/#evidence-stops) | [6](/vendors/check-point/#evidence-layers) | [5](/vendors/check-point/#evidence-latency) | [6](/vendors/check-point/#evidence-deploy) | [4](/vendors/check-point/#evidence-response) | [6](/vendors/check-point/#evidence-transparency) | [7](/vendors/check-point/#evidence-proof) | | 10th | [CrowdStrike](/vendors/crowdstrike/) | 52.0 | [6](/vendors/crowdstrike/#evidence-stops) | [7](/vendors/crowdstrike/#evidence-layers) | [0](/vendors/crowdstrike/#evidence-latency) | [6](/vendors/crowdstrike/#evidence-deploy) | [7](/vendors/crowdstrike/#evidence-response) | [4](/vendors/crowdstrike/#evidence-transparency) | [7](/vendors/crowdstrike/#evidence-proof) | | 11th | [Aim Security](/vendors/aim-security/) | 50.5 | [6](/vendors/aim-security/#evidence-stops) | [7](/vendors/aim-security/#evidence-layers) | [0](/vendors/aim-security/#evidence-latency) | [6](/vendors/aim-security/#evidence-deploy) | [5](/vendors/aim-security/#evidence-response) | [3](/vendors/aim-security/#evidence-transparency) | [8](/vendors/aim-security/#evidence-proof) | | 12th | [Zenity](/vendors/zenity/) | 50.0 | [7](/vendors/zenity/#evidence-stops) | [6](/vendors/zenity/#evidence-layers) | [0](/vendors/zenity/#evidence-latency) | [4](/vendors/zenity/#evidence-deploy) | [6](/vendors/zenity/#evidence-response) | [4](/vendors/zenity/#evidence-transparency) | [8](/vendors/zenity/#evidence-proof) | | 13th | [Cloudflare](/vendors/cloudflare/) | 46.0 | [5](/vendors/cloudflare/#evidence-stops) | [5](/vendors/cloudflare/#evidence-layers) | [0](/vendors/cloudflare/#evidence-latency) | [6](/vendors/cloudflare/#evidence-deploy) | [6](/vendors/cloudflare/#evidence-response) | [6](/vendors/cloudflare/#evidence-transparency) | [6](/vendors/cloudflare/#evidence-proof) | | 14th | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | 45.5 | [5](/vendors/sentinelone/#evidence-stops) | [6](/vendors/sentinelone/#evidence-layers) | [0](/vendors/sentinelone/#evidence-latency) | [6](/vendors/sentinelone/#evidence-deploy) | [4](/vendors/sentinelone/#evidence-response) | [4](/vendors/sentinelone/#evidence-transparency) | [7](/vendors/sentinelone/#evidence-proof) | | 15th | [Microsoft](/vendors/microsoft/) | 45.0 | [4](/vendors/microsoft/#evidence-stops) | [6](/vendors/microsoft/#evidence-layers) | [0](/vendors/microsoft/#evidence-latency) | [6](/vendors/microsoft/#evidence-deploy) | [6](/vendors/microsoft/#evidence-response) | [6](/vendors/microsoft/#evidence-transparency) | [6](/vendors/microsoft/#evidence-proof) | | 16th | [Noma Security](/vendors/noma-security/) | 41.5 | [5](/vendors/noma-security/#evidence-stops) | [6](/vendors/noma-security/#evidence-layers) | [0](/vendors/noma-security/#evidence-latency) | [4](/vendors/noma-security/#evidence-deploy) | [4](/vendors/noma-security/#evidence-response) | [3](/vendors/noma-security/#evidence-transparency) | [7](/vendors/noma-security/#evidence-proof) | ## Who should pick what With no particular situation, the published ranking applies and HiddenLayer comes first. For a specific situation the weights change, and so can the order: - **[Developers run coding agents with shell, file and package access](/best-for/coding-agents/)**: pick HiddenLayer (74.5). Bay is 3rd at 67.5. A coding agent acts on the machine, so stopping named actions there before they complete weighs most. - **[A company shipping a customer-facing LLM application](/best-for/customer-llm-app/)**: pick HiddenLayer (68.0). Bay is joint 3rd at 63.5. An inline control in a customer path must be fast and fail predictably, so latency and failure mode weigh more. - **[Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/)**: pick HiddenLayer (77.5). Bay is joint 4th at 67.0. Agents in SaaS act far from the endpoint, so coverage across layers weighs as much as what is stopped. - **[A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/)**: pick HiddenLayer (77.0). Bay is joint 3rd at 66.5. Proving control means an audit trail and a response path, so detection and response weighs most after what is stopped. - **[A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/)**: pick HiddenLayer (69.5). Bay is 2nd at 69.0. A lean team with a deadline needs the lightest deployment that still stops actions. ## What would change this ranking **Top pair.** Cisco rising from 7 to 9 on action types stopped before they complete (weight 30, so each point moves a total by 3.0) would put Cisco above HiddenLayer; the gap is 4.5 points. **Bay's next place.** Bay rising from 8 to 9 on action types stopped before they complete (weight 30, so each point moves a total by 3.0) would put Bay above Lasso Security; the gap is 1.0 points. A score changes only when a sourced fact changes, and the weights were fixed before the scores. The [method page](/method/) lists the weights, anchors and rules. ## What does an AI runtime security tool do? An AI runtime security tool watches an AI system while it works and stops a harmful step before the step completes. The step can be a prompt, a model answer, an agent tool call or an action on a machine, such as a shell command or a file read. This site compares tools by the actions they stop, not by feature counts. Every score comes from a vendor evidence file with a quote and a read date, and every gap is written as none found on 16 Nov 2026. ## How do you use the action picker? Choose an action in the picker above: shell command, file access, data paste, package install, network call, credential access, prompt injection or harmful model output. The page then lists which of the 16 tools document a stop for that action, which document part of one, and which have none found. A stop counts only where the vendor names the action and the block happens before it completes. A detect-only or alert-only mention is marked partial. With scripts off, the same matrix is a plain table on the Runtime Action Type Matrix page. ## Which question should you ask first? Ask where the harm happens. A customer-facing chatbot needs a control in the request path with a published failure mode. A coding agent needs a control on the developer machine. Agents inside SaaS need coverage across several layers. The five situation pages re-weight the same scores for those cases. Weights were committed on 16 Nov 2026 before any evidence or score existed, so no situation page can change a score. ## Latest news, guides and data - [Cisco open sources Project Quarry agent detections](/news/cisco-project-quarry-agent-detections/) (2026-10-28) - [Cisco DefenseClaw pairs with NVIDIA OpenShell sandbox](/news/cisco-defenseclaw-openshell-coding-agents/) (2026-10-21) - [Bedrock Guardrails grounding checks add Standard tier](/news/bedrock-guardrails-grounding-standard-tier/) (2026-10-19) - [AI Runtime Security Explained](/guides/ai-runtime-security-explained/) - [AI Runtime Security vs Guardrails vs Red Teaming](/guides/ai-runtime-security-vs-guardrails-vs-red-teaming/) - [AIDR vs EDR vs XDR: What Each Watches](/guides/aidr-vs-edr-vs-xdr/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/), with [CSV and JSON downloads](/datasets/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/), with [CSV and JSON downloads](/datasets/enforcement-point-map/) ## Related pages - [Full ranking with re-rank sliders](/ranking/) - [How the scores and weights work](/method/) - [Vendor profiles with evidence](/vendors/) - [Best tool by situation](/best-for/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) - [Guides](/guides/) - [Frequently asked questions](/faq/) --- # AI runtime security tools ranking URL: https://airuntimesecuritytools.com/ranking/ Updated: 2026-11-16 HiddenLayer ranks first among 16 AI runtime security tools with 71.0 out of 100 on 16 Nov 2026. Bay is joint 4th with 62.5. Totals are weighted sums of seven criteria, led by the actions a tool stops before they complete, and each total traces to dated facts. ## The short answer HiddenLayer leads the site ranking at 71.0 of 100. Bay is joint 4th at 62.5. HiddenLayer is ahead mainly on layers covered (9 against 6, worth 6.0 points at weight 20). **Site ranking: scores out of 10 per criterion, totals out of 100** | Rank | Vendor | Total (of 100) | Action types stopped before they complete (weight 30) | Layers covered (weight 20) | Latency and failure mode (weight 15) | Deployment (weight 15) | Detection and response (weight 5) | Transparency (weight 10) | Proof (weight 5) | |---|---|---|---|---|---|---|---|---|---| | 1st | [HiddenLayer](/vendors/hiddenlayer/) | 71.0 | [8](/vendors/hiddenlayer/#evidence-stops) | [9](/vendors/hiddenlayer/#evidence-layers) | [4](/vendors/hiddenlayer/#evidence-latency) | [6](/vendors/hiddenlayer/#evidence-deploy) | [9](/vendors/hiddenlayer/#evidence-response) | [6](/vendors/hiddenlayer/#evidence-transparency) | [7](/vendors/hiddenlayer/#evidence-proof) | | 2nd | [Cisco](/vendors/cisco/) | 66.5 | [7](/vendors/cisco/#evidence-stops) | [8](/vendors/cisco/#evidence-layers) | [5](/vendors/cisco/#evidence-latency) | [7](/vendors/cisco/#evidence-deploy) | [8](/vendors/cisco/#evidence-response) | [5](/vendors/cisco/#evidence-transparency) | [5](/vendors/cisco/#evidence-proof) | | 3rd | [Lasso Security](/vendors/lasso-security/) | 63.5 | [7](/vendors/lasso-security/#evidence-stops) | [7](/vendors/lasso-security/#evidence-layers) | [5](/vendors/lasso-security/#evidence-latency) | [7](/vendors/lasso-security/#evidence-deploy) | [7](/vendors/lasso-security/#evidence-response) | [3](/vendors/lasso-security/#evidence-transparency) | [8](/vendors/lasso-security/#evidence-proof) | | joint 4th | [Bay](/vendors/bay/) | 62.5 | [8](/vendors/bay/#evidence-stops) | [6](/vendors/bay/#evidence-layers) | [5](/vendors/bay/#evidence-latency) | [8](/vendors/bay/#evidence-deploy) | [7](/vendors/bay/#evidence-response) | [0](/vendors/bay/#evidence-transparency) | [7](/vendors/bay/#evidence-proof) | | joint 4th | [Operant AI](/vendors/operant-ai/) | 62.5 | [7](/vendors/operant-ai/#evidence-stops) | [7](/vendors/operant-ai/#evidence-layers) | [4](/vendors/operant-ai/#evidence-latency) | [7](/vendors/operant-ai/#evidence-deploy) | [7](/vendors/operant-ai/#evidence-response) | [4](/vendors/operant-ai/#evidence-transparency) | [7](/vendors/operant-ai/#evidence-proof) | | joint 4th | [Straiker](/vendors/straiker/) | 62.5 | [7](/vendors/straiker/#evidence-stops) | [7](/vendors/straiker/#evidence-layers) | [5](/vendors/straiker/#evidence-latency) | [7](/vendors/straiker/#evidence-deploy) | [5](/vendors/straiker/#evidence-response) | [3](/vendors/straiker/#evidence-transparency) | [8](/vendors/straiker/#evidence-proof) | | 7th | [Palo Alto Networks](/vendors/palo-alto-networks/) | 58.5 | [6](/vendors/palo-alto-networks/#evidence-stops) | [7](/vendors/palo-alto-networks/#evidence-layers) | [4](/vendors/palo-alto-networks/#evidence-latency) | [6](/vendors/palo-alto-networks/#evidence-deploy) | [4](/vendors/palo-alto-networks/#evidence-response) | [6](/vendors/palo-alto-networks/#evidence-transparency) | [7](/vendors/palo-alto-networks/#evidence-proof) | | 8th | [Pillar Security](/vendors/pillar-security/) | 55.5 | [7](/vendors/pillar-security/#evidence-stops) | [7](/vendors/pillar-security/#evidence-layers) | [0](/vendors/pillar-security/#evidence-latency) | [6](/vendors/pillar-security/#evidence-deploy) | [7](/vendors/pillar-security/#evidence-response) | [4](/vendors/pillar-security/#evidence-transparency) | [8](/vendors/pillar-security/#evidence-proof) | | 9th | [Check Point](/vendors/check-point/) | 55.0 | [5](/vendors/check-point/#evidence-stops) | [6](/vendors/check-point/#evidence-layers) | [5](/vendors/check-point/#evidence-latency) | [6](/vendors/check-point/#evidence-deploy) | [4](/vendors/check-point/#evidence-response) | [6](/vendors/check-point/#evidence-transparency) | [7](/vendors/check-point/#evidence-proof) | | 10th | [CrowdStrike](/vendors/crowdstrike/) | 52.0 | [6](/vendors/crowdstrike/#evidence-stops) | [7](/vendors/crowdstrike/#evidence-layers) | [0](/vendors/crowdstrike/#evidence-latency) | [6](/vendors/crowdstrike/#evidence-deploy) | [7](/vendors/crowdstrike/#evidence-response) | [4](/vendors/crowdstrike/#evidence-transparency) | [7](/vendors/crowdstrike/#evidence-proof) | | 11th | [Aim Security](/vendors/aim-security/) | 50.5 | [6](/vendors/aim-security/#evidence-stops) | [7](/vendors/aim-security/#evidence-layers) | [0](/vendors/aim-security/#evidence-latency) | [6](/vendors/aim-security/#evidence-deploy) | [5](/vendors/aim-security/#evidence-response) | [3](/vendors/aim-security/#evidence-transparency) | [8](/vendors/aim-security/#evidence-proof) | | 12th | [Zenity](/vendors/zenity/) | 50.0 | [7](/vendors/zenity/#evidence-stops) | [6](/vendors/zenity/#evidence-layers) | [0](/vendors/zenity/#evidence-latency) | [4](/vendors/zenity/#evidence-deploy) | [6](/vendors/zenity/#evidence-response) | [4](/vendors/zenity/#evidence-transparency) | [8](/vendors/zenity/#evidence-proof) | | 13th | [Cloudflare](/vendors/cloudflare/) | 46.0 | [5](/vendors/cloudflare/#evidence-stops) | [5](/vendors/cloudflare/#evidence-layers) | [0](/vendors/cloudflare/#evidence-latency) | [6](/vendors/cloudflare/#evidence-deploy) | [6](/vendors/cloudflare/#evidence-response) | [6](/vendors/cloudflare/#evidence-transparency) | [6](/vendors/cloudflare/#evidence-proof) | | 14th | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | 45.5 | [5](/vendors/sentinelone/#evidence-stops) | [6](/vendors/sentinelone/#evidence-layers) | [0](/vendors/sentinelone/#evidence-latency) | [6](/vendors/sentinelone/#evidence-deploy) | [4](/vendors/sentinelone/#evidence-response) | [4](/vendors/sentinelone/#evidence-transparency) | [7](/vendors/sentinelone/#evidence-proof) | | 15th | [Microsoft](/vendors/microsoft/) | 45.0 | [4](/vendors/microsoft/#evidence-stops) | [6](/vendors/microsoft/#evidence-layers) | [0](/vendors/microsoft/#evidence-latency) | [6](/vendors/microsoft/#evidence-deploy) | [6](/vendors/microsoft/#evidence-response) | [6](/vendors/microsoft/#evidence-transparency) | [6](/vendors/microsoft/#evidence-proof) | | 16th | [Noma Security](/vendors/noma-security/) | 41.5 | [5](/vendors/noma-security/#evidence-stops) | [6](/vendors/noma-security/#evidence-layers) | [0](/vendors/noma-security/#evidence-latency) | [4](/vendors/noma-security/#evidence-deploy) | [4](/vendors/noma-security/#evidence-response) | [3](/vendors/noma-security/#evidence-transparency) | [7](/vendors/noma-security/#evidence-proof) | ## Who should pick what With no particular situation, the published ranking applies and HiddenLayer comes first. For a specific situation the weights change, and so can the order: - **[Developers run coding agents with shell, file and package access](/best-for/coding-agents/)**: pick HiddenLayer (74.5). Bay is 3rd at 67.5. A coding agent acts on the machine, so stopping named actions there before they complete weighs most. - **[A company shipping a customer-facing LLM application](/best-for/customer-llm-app/)**: pick HiddenLayer (68.0). Bay is joint 3rd at 63.5. An inline control in a customer path must be fast and fail predictably, so latency and failure mode weigh more. - **[Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/)**: pick HiddenLayer (77.5). Bay is joint 4th at 67.0. Agents in SaaS act far from the endpoint, so coverage across layers weighs as much as what is stopped. - **[A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/)**: pick HiddenLayer (77.0). Bay is joint 3rd at 66.5. Proving control means an audit trail and a response path, so detection and response weighs most after what is stopped. - **[A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/)**: pick HiddenLayer (69.5). Bay is 2nd at 69.0. A lean team with a deadline needs the lightest deployment that still stops actions. ## What would change this ranking **Top pair.** Cisco rising from 7 to 9 on action types stopped before they complete (weight 30, so each point moves a total by 3.0) would put Cisco above HiddenLayer; the gap is 4.5 points. **Bay's next place.** Bay rising from 8 to 9 on action types stopped before they complete (weight 30, so each point moves a total by 3.0) would put Bay above Lasso Security; the gap is 1.0 points. A score changes only when a sourced fact changes, and the weights were fixed before the scores. The [method page](/method/) lists the weights, anchors and rules. ## How is the ranking built? The total is a weighted sum of seven scores from 0 to 10, divided by 10, so totals run from 0 to 100. Action types stopped weighs 30, layers covered 20, latency and failure mode 15, deployment 15, transparency 10, detection and response 5 and proof 5. A vendor claim alone earns at most 7 on any criterion. Only a second source, such as third-party coverage or public documentation observed directly, can earn 8 to 10. ## What does the ranking not measure? The ranking does not measure detection accuracy, false positive rates or bypass resistance, because no vendor publishes those figures in a form that can be compared. Third-party testing of that kind is the largest gap in this market. Latency and failure mode are scored only as published. A vendor with no published figure scores low on that criterion even if its product is fast in practice. ## Why do totals differ by situation? Each situation page keeps the same 0 to 10 scores and changes only the weights, for example giving latency 25 for a customer-facing app. The leader can change because the question changed, not because a score moved. Re-rank sliders let you set your own weights. A link restores the published ones. ## Related pages - [Method: weights, anchors and rules](/method/) - [Rankings by situation](/best-for/) - [Vendor profiles](/vendors/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) - [Guides for buyers](/guides/) - [Glossary](/glossary/) --- # How we score AI runtime security tools URL: https://airuntimesecuritytools.com/method/ Updated: 2026-11-16 One research plan covered all 16 vendors on 16 Nov 2026 and produced 1,469 facts across 32 vendor files, each checked against a quote. One scorer scored blind, a second blind reviewer checked, and weights were committed before any evidence or score existed. ## The weights The weights were fixed on 2026-11-16, before any score existed, and have not changed since. **Criteria and weights (out of 100)** | Criterion | What it measures | Weight | |---|---|---| | Action types stopped before they complete | Which harmful actions it stops before they complete, named: shell command, file read or write, data paste or upload, package install, network call, credential access, unsafe tool call, prompt injection or harmful model output | 30 | | Layers covered | Which layers it enforces at, as documented: model (prompts and outputs), application (an AI app's requests), agent (tool calls and plans), endpoint (actions on the machine) | 20 | | Latency and failure mode | Published latency of an inline decision and the published behavior when the control fails (fail open or fail closed, timeouts) | 15 | | Deployment | How it deploys (API, SDK, proxy or gateway, agent, agentless), what it needs, time to deploy as published | 15 | | Detection and response | Alerts, an audit trail of blocked and allowed actions, investigation, and export to SIEM or SOAR | 5 | | Transparency | Public documentation readable without sign-in, a published price, published research | 10 | | Proof | Named customers, funding, independent analyst coverage | 5 | ## Weights by situation **Weights used for each situation view** | Situation | Action types stopped before they complete | Layers covered | Latency and failure mode | Deployment | Detection and response | Transparency | Proof | |---|---|---|---|---|---|---|---| | [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) | 35 | 20 | 10 | 15 | 10 | 5 | 5 | | [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) | 25 | 20 | 25 | 15 | 5 | 5 | 5 | | [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) | 30 | 30 | 5 | 15 | 10 | 5 | 5 | | [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) | 25 | 15 | 10 | 10 | 30 | 5 | 5 | | [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) | 25 | 15 | 10 | 35 | 5 | 5 | 5 | ## Score anchors - **0**: none found on the read date - **1-3**: mentioned, partial, or limited to one surface, one layer or one action type - **4-6**: documented and working, with gaps the documentation shows - **7**: the most a vendor claim or vendor documentation alone can earn: detailed, specific, covering the whole criterion - **8-10**: detailed and confirmed by a second source (third-party write-up, or observed directly in public documentation, code, a marketplace listing or a release record) ## Rules - Every vendor is scored 0 to 10 on every criterion, whole numbers, from the evidence file only, by a scorer that works from coded packets with every vendor name removed. - A vendor claim scores at most 7 until a second source confirms it (kind third-party or observed). - None found scores 0 and is written "none found on the read date". - Breadth facts (module counts, integration counts, headcount, funding) inform Proof only, which weighs 5, so they carry at most 10 percent of any ranking. - The total is the weighted sum of the scores divided by 10 (0 to 100 scale, published to one decimal). - Equal totals print as a joint rank (for example joint 2nd); inside a joint rank vendors are listed alphabetically for display only. - Once the first score exists, no weight and no criterion changes. A score changes only when a sourced fact changes, logged with the fact. ## What research was done? A single research plan was applied to every vendor on 16 Nov 2026. The same questions were asked of each: which actions are stopped, at which layers, with what published latency and failure mode, how it deploys, what it records, what is public and who backs it. The result is 1,469 facts across 32 vendor files, because the files also serve other comparisons; the 16 files for this site hold 896 of them. Each fact holds a claim, a quote, a source URL, a read date and a kind: vendor, third-party, observed or none-found. Every quote was checked against the page it came from. ## How does blind scoring work? One scorer scored every vendor from the evidence file only, working from coded packets with every vendor name removed, so the scorer did not know which vendor was which. One blind reviewer then checked each score against its cited facts. Disagreements were settled by returning to the quote. Scores are whole numbers from 0 to 10. A score changes only when a sourced fact changes, and the change is logged with the fact. ## Why are weights fixed first? The weights, the five situation weightings and the scoring anchors were committed on 16 Nov 2026 before any vendor evidence or score existed. After the first score, no weight and no criterion changes. Breadth counts such as module numbers, integrations, headcount and funding inform only the proof criterion, which weighs 5, so they carry at most 10 percent of any ranking. ## How are latency and failure mode scored? Only as published. A millisecond figure on a vendor page, a documented timeout or a stated fail-open or fail-closed default earns credit. A figure seen only in a demo panel is treated as a mention. Silence scores 0 and is written as none found on 16 Nov 2026. This favors vendors that document their control, which is the point: a buyer cannot plan around a failure mode nobody has written down. ## How are errors corrected? Send a correction with a source to editor@airuntimesecuritytools.com. A confirmed fact change updates the fact, the score and the changelog on the same day, and the old and new values stay visible. Vendors may supply public documentation at any time. A claim without a public source stays at the claim ceiling of 7. ## Changelog - 2026-11-16: weights and criteria fixed. - 2026-11-16: evidence read for every vendor by one research plan. - 2026-11-16: scores published (blind: one scorer, then one blind reviewer). - 2026-11-16: this site built from the committed scores. ## Related pages - [The ranking these weights produce](/ranking/) - [Vendor profiles with evidence per score](/vendors/) - [Editorial policy](/editorial-policy/) - [Corrections](/corrections/) - [Glossary](/glossary/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # AI runtime security vendors URL: https://airuntimesecuritytools.com/vendors/ Updated: 2026-11-16 Sixteen vendor profiles list what each AI runtime security tool stops before it completes, the layers it enforces at, published latency, deployment path and proof, with scores out of 10 and dated facts. Every profile states one strength and one gap from the 16 Nov 2026 evidence. ## All 16 vendors **Vendors in the ranking, with their rank and total** | Vendor | Rank | Total (of 100) | Product assessed | |---|---|---|---| | [Aim Security](/vendors/aim-security/) | 11th | 50.5 | Cato AI Security: AI-Firewall for Applications, AI Security for Agents and Cato AI Endpoint | | [Bay](/vendors/bay/) | joint 4th | 62.5 | Bay Agent Runtime Enforcement (Bay Enforcer, Lighthouse policy engine) | | [Check Point](/vendors/check-point/) | 9th | 55.0 | Check Point AI Guardrails (Guard API) and AI Agent Security runtime protection | | [Cisco](/vendors/cisco/) | 2nd | 66.5 | Cisco AI Defense runtime protection, Python SDK and Kong plugin inspection, and DefenseClaw endpoint hooks | | [Cloudflare](/vendors/cloudflare/) | 13th | 46.0 | AI Security for Apps (formerly Firewall for AI), AI Gateway Guardrails and DLP, and Cloudflare One MCP server portals | | [CrowdStrike](/vendors/crowdstrike/) | 10th | 52.0 | Falcon Guardian (sensor, agent hooks, SafeMind detection), Falcon AI Gateway, Kill Switch, and Falcon sensor package protection | | [HiddenLayer](/vendors/hiddenlayer/) | 1st | 71.0 | AI Runtime Security (AIDR and Agentic Runtime Security) plus Agent Harness Security for coding agents | | [Lasso Security](/vendors/lasso-security/) | 3rd | 63.5 | Lasso AI Detection & Response with Intent Security Engine and LEAP/RAPID guardrails | | [Microsoft](/vendors/microsoft/) | 15th | 45.0 | Defender for Endpoint AI agent runtime protection (preview), Azure AI Content Safety Prompt Shields, Defender for Cloud AI threat protection, Agent 365 real-time protection | | [Noma Security](/vendors/noma-security/) | 16th | 41.5 | Noma AI-DR (AI Detection and Response) with Agent Access Control | | [Operant AI](/vendors/operant-ai/) | joint 4th | 62.5 | Semantic Firewall, Agent Protector (with CodeInjectionGuard and ScopeGuard), Endpoint Protector and AI Gatekeeper | | [Palo Alto Networks](/vendors/palo-alto-networks/) | 7th | 58.5 | Prisma AIRS AI Runtime Security (API Intercept, Network Intercept, AI Gateway) | | [Pillar Security](/vendors/pillar-security/) | 8th | 55.5 | Runtime Guardrails and Agentic Endpoint (Runtime Guardian) | | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | 14th | 45.5 | Prompt Security AI Application Security and AI firewall, plus agent and endpoint enforcement (SentinelOne) | | [Straiker](/vendors/straiker/) | joint 4th | 62.5 | Defend AI (runtime security and guardrails) with the Agentic Kill Switch | | [Zenity](/vendors/zenity/) | 12th | 50.0 | Zenity Boundaries (runtime enforcement engine) and AI Detection & Response (AIDR), with native agent hooks and the Zenity MCP gateway for coding agents | ## Related pages - [Ranking of all vendors](/ranking/) - [How vendors are scored](/method/) - [Best vendor by situation](/best-for/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # Aim Security AI runtime security URL: https://airuntimesecuritytools.com/vendors/aim-security/ Updated: 2026-11-16 Aim Security is now part of Cato Networks, and its runtime product is an AI firewall. Cato says it detects jailbreaks, prompt injection and off-policy responses and blocks or redacts them inline, with unapproved agent tool calls and data paste also stopped before execution. Deployment options include a device-managed endpoint client, inline network inspection of AI traffic, a proxy, gateway or API, and on-prem. Customers are quoted by name, and an IDC analyst is cited. Gaps from 16 Nov 2026: a latency figure, fail-open or fail-closed behavior, blocking of shell, file, package, network and credential actions, readable product docs and a price are none found. ## Strength and weakness for AI Runtime Security Tools **Strength:** Inline inspection blocks prompt injection, jailbreaks, unsafe outputs and unapproved agent tool calls across model, application, agent and endpoint layers (aim-security-59, -41, -63). **Weakness:** None found on 2026-11-16 for a published latency figure, fail-open or fail-closed behavior, or blocking of shell, file, package, network or credential actions (aim-security-66, -62). ## Scores and evidence Aim Security is 11th in the overall ranking at 50.5 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Aim Security: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Action types stopped before they complete | 6 of 10 | 30 | The vendor claims stops for data paste, unapproved tool calls, prompt injection and unsafe outputs before execution. Shell, file, package, network and credential blocking: none found. | [aim-security-23](#fact-aim-security-23), [aim-security-41](#fact-aim-security-41), [aim-security-59](#fact-aim-security-59), [aim-security-61](#fact-aim-security-61), [aim-security-52](#fact-aim-security-52), [aim-security-62](#fact-aim-security-62) | | Layers covered | 7 of 10 | 20 | The vendor claims endpoint, agent, application and model coverage. The supporting release is the vendor's own wire release. | [aim-security-16](#fact-aim-security-16), [aim-security-41](#fact-aim-security-41), [aim-security-49](#fact-aim-security-49), [aim-security-71](#fact-aim-security-71), [aim-security-53](#fact-aim-security-53) | | Latency and failure mode | 0 of 10 | 15 | The vendor claims millisecond checks with no figure. Fail-open or fail-closed behavior: none found. | [aim-security-66](#fact-aim-security-66), [aim-security-64](#fact-aim-security-64) | | Deployment | 6 of 10 | 15 | Endpoint client through device management, proxy, gateway or API, and on-prem Outposts, live in weeks. Supported OS list: none found. | [aim-security-19](#fact-aim-security-19), [aim-security-20](#fact-aim-security-20), [aim-security-49](#fact-aim-security-49), [aim-security-70](#fact-aim-security-70), [aim-security-21](#fact-aim-security-21) | | Detection and response | 5 of 10 | 5 | The vendor claims logging, a dashboard of blocked and monitored results, and audit trails. Named SIEM or SOAR integrations: none found. | [aim-security-68](#fact-aim-security-68), [aim-security-69](#fact-aim-security-69), [aim-security-56](#fact-aim-security-56), [aim-security-22](#fact-aim-security-22), [aim-security-57](#fact-aim-security-57) | | Transparency | 3 of 10 | 10 | Research and white papers are on the parent's blog. Readable product docs and a published price: none found. | [aim-security-32](#fact-aim-security-32), [aim-security-33](#fact-aim-security-33), [aim-security-06](#fact-aim-security-06), [aim-security-30](#fact-aim-security-30), [aim-security-31](#fact-aim-security-31) | | Proof | 8 of 10 | 5 | Third-party sources cover the parent's funding and the acquired company's research, and the vendor quotes named customers and an IDC analyst. | [aim-security-07](#fact-aim-security-07), [aim-security-08](#fact-aim-security-08), [aim-security-09](#fact-aim-security-09), [aim-security-34](#fact-aim-security-34), [aim-security-35](#fact-aim-security-35), [aim-security-36](#fact-aim-security-36) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 11th | 50.5 | | [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) | joint 10th | 54.5 | | [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) | 11th | 46.0 | | [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) | 10th | 58.5 | | [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) | 11th | 52.0 | | [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) | 11th | 54.5 | ## Product, deployment, price and owner - Product assessed: Cato AI Security: AI-Firewall for Applications, AI Security for Agents and Cato AI Endpoint - Deployment: Lightweight endpoint client pushed through existing device management, plus inline network inspection of AI traffic; proxy, AI gateway or API integration for applications; on-prem or own-cloud Outposts option - Time to deploy: Cato says AI protection goes live in weeks; no operating system list found on 2026-11-16 - Price: none found on 2026-11-16; the product page offers a demo and the /pricing/ page returned 404 - Owner: Cato Networks - Website: [aim.security](https://www.aim.security) ## Facts and sources - **aim-security-01** (vendor): Cato Networks says in its own blog dated 2025-09-03 that it acquired Aim Security, and that Aim would stay available as a standalone product before converging into the Cato platform by early 2026.Quote: "In the near term, Aim will remain available as a standalone product. By early 2026, it will converge into the Cato platform"Source: [catonetworks.com](https://www.catonetworks.com/blog/securing-ai-transformation-why-cato-acquired-aim-security/), read 2026-11-16. - **aim-security-02** (third-party): SecurityWeek reported on 2025-09-03 that Cato Networks acquired Aim Security, which was founded in 2022 and came out of stealth in January 2024.Quote: "Founded in 2022 to help organizations with the secure deployment of generative-AI utilities, Aim emerged from stealth in January 2024."Source: [securityweek.com](https://www.securityweek.com/cato-networks-acquires-ai-security-firm-aim-security/), read 2026-11-16. - **aim-security-03** (third-party): In a PRNewswire release dated 2026-07-27, Cato says its Cato AI Security product combines capabilities from its Aim Security acquisition and was introduced to market in March 2026.Quote: "Cato AI Security, which converges advanced AI governance and protection capabilities from Cato's Aim Security acquisition into a fully integrated solution introduced to market in March 2026"Source: [prnewswire.com](https://www.prnewswire.com/news-releases/cato-networks-exceeds-415m-in-arr-fueled-by-42-yoy-growth-and-enterprise-demand-302833667.html), read 2026-11-16. - **aim-security-04** (third-party): Wikipedia states that Cato Networks announced its first M&A deal, the acquisition of Aim Security, in September 2025.Quote: "Cato Networks announced its first-ever M&A deal with the acquisition of Aim Security, an AI security startup based in Israel"Source: [en.wikipedia.org](https://en.wikipedia.org/wiki/Cato_Networks), read 2026-11-16. - **aim-security-05** (vendor): Cato posts a Bloomberg summary dated 2025-09-07 saying Cato agreed to acquire Aim Security, Cato's first acquisition.Quote: "Israeli cybersecurity company Cato Networks has agreed to acquire AI startup Aim Security, marking the firm’s first acquisition."Source: [catonetworks.com](https://www.catonetworks.com/news/bloomberg-cato-networks-to-buy-aim-security-in-bid-to-make-ai-more-secure/), read 2026-11-16. - **aim-security-06** (none-found): None found on 2026-11-16: https://www.aim.security would not load (HTTP 403 Cloudflare block, also blocked via the r.jina.ai reader and /sitemap.xml), so no Aim-owned page could be read; ownership and product facts come from Cato Networks pages and third-party reports.Source: [aim.security](https://www.aim.security), read 2026-11-16. - **aim-security-07** (third-party): Wikipedia reports Cato Networks raised $50 million more in a Series G, bringing the total to $409 million at an estimated valuation above $4.8 billion.Quote: "it had raised another $50 million in its Series G round, bringing the total to $409 million at an estimated valuation of over $4.8 billion"Source: [en.wikipedia.org](https://en.wikipedia.org/wiki/Cato_Networks), read 2026-11-16. - **aim-security-08** (third-party): SecurityWeek reports Aim raised $10 million in seed funding and $18 million more in June 2025 before the acquisition, and that its researchers found the EchoLeak Microsoft 365 Copilot vulnerability (CVE-2025-32711).Quote: "Aim’s research team recently uncovered the first reported CVE of a zero-click AI vulnerability in Microsoft 365 Copilot"Source: [securityweek.com](https://www.securityweek.com/cato-networks-acquires-ai-security-firm-aim-security/), read 2026-11-16. - **aim-security-09** (third-party): Cato says on 2026-07-27 that it passed $415 million in ARR with 42% year-over-year growth and serves more than 4,800 customers.Quote: "surpassed $415 million in ARR, with 42% year-over-year (YoY) growth"Source: [prnewswire.com](https://www.prnewswire.com/news-releases/cato-networks-exceeds-415m-in-arr-fueled-by-42-yoy-growth-and-enterprise-demand-302833667.html), read 2026-11-16. - **aim-security-16** (vendor): Cato says AI Endpoint uses a lightweight client for browser AI and desktop AI apps, and also inspects browser and app AI traffic inline on the network with the same engine.Quote: "Browser AI and AI apps are also inspected inline on the network with the same AI engine, so endpoint and network AI events share one dashboard."Source: [catonetworks.com](https://www.catonetworks.com/platform/cato-ai-endpoint/), read 2026-11-16. - **aim-security-19** (vendor): Cato says AI Endpoint is pushed out with existing endpoint management tools, runs beside the EDR, and AI protection goes live in weeks.Quote: "Existing SSE, VPN and EDR keep doing their jobs, so AI protection goes live in weeks."Source: [catonetworks.com](https://www.catonetworks.com/platform/cato-ai-endpoint/), read 2026-11-16. - **aim-security-20** (vendor): Cato says the AI Endpoint client is deployed through device management like any other app, and finding agents needs no network change.Quote: "The Cato AI Endpoint client goes out through device management like any other app."Source: [catonetworks.com](https://www.catonetworks.com/blog/cato-ai-endpoint-secures-browser-ai-desktop-ai-and-coding-agents-on-every-endpoint/), read 2026-11-16. - **aim-security-21** (none-found): None found on 2026-11-16: the Cato AI Endpoint page and launch blog do not list supported operating systems or browsers, or a time to first inventory in hours or days.Source: [catonetworks.com](https://www.catonetworks.com/platform/cato-ai-endpoint/), read 2026-11-16. - **aim-security-22** (vendor): Cato lists enforcement actions: block or redact sensitive data, redirect users to approved tools, and log violations for evidence.Quote: "Automatically block or redact sensitive data"Source: [catonetworks.com](https://www.catonetworks.com/platform/ai-security-for-end-users/), read 2026-11-16. - **aim-security-23** (vendor): Cato says that when a prompt is sent from a desktop AI app, AI Endpoint checks it and stops a pasted customer list from reaching the AI service, with a pop-up telling the user.Quote: "Cato AI Endpoint checks the prompt and stops the customer list from reaching the AI service."Source: [catonetworks.com](https://www.catonetworks.com/platform/cato-ai-endpoint/), read 2026-11-16. - **aim-security-27** (vendor): Cato says AI Endpoint shows each agent by user and endpoint with its MCP servers and skills, listing agents with risky settings first.Quote: "Agents with risky settings, such as unrestricted access to the command line, show up first, so the team knows where to start."Source: [catonetworks.com](https://www.catonetworks.com/platform/cato-ai-endpoint/), read 2026-11-16. - **aim-security-30** (none-found): None found on 2026-11-16: no published price for Cato AI Security or Cato AI Endpoint on the product page, which offers "Get a product demo"; https://www.catonetworks.com/pricing/ returned HTTP 404.Source: [catonetworks.com](https://www.catonetworks.com/platform/cato-ai-endpoint/), read 2026-11-16. - **aim-security-31** (none-found): None found on 2026-11-16: the Cato support and documentation site at support.catonetworks.com returned HTTP 403 to an automated request, so public product documentation could not be read.Source: [support.catonetworks.com](https://support.catonetworks.com/), read 2026-11-16. - **aim-security-32** (vendor): Cato publishes threat research on its public blog; a 2025-04-15 Cato CTRL post covers risks of the Model Context Protocol and security practices.Quote: "As MCP-enabled tools interact with sensitive data, they increase the risk of unauthorized access."Source: [catonetworks.com](https://www.catonetworks.com/blog/cato-ctrl-exploiting-model-context-protocol-mcp/), read 2026-11-16. - **aim-security-33** (vendor): Cato offers public product pages, a white paper and a webinar on agentic AI security.Quote: "Securing agentic AI white paper"Source: [catonetworks.com](https://www.catonetworks.com/platform/agentic-ai-security/), read 2026-11-16. - **aim-security-34** (vendor): Cato publishes a quote from Nithin Reddy, VP Security at Dayforce, who says Dayforce deployed Cato AI extensions and put Cato AI security inline.Quote: "We prioritized the AI our employees use, so we deployed Cato AI extensions for visibility and to prevent prompt injection and data leakage."Source: [catonetworks.com](https://www.catonetworks.com/platform/ai-security-for-end-users/), read 2026-11-16. - **aim-security-35** (vendor): Cato publishes a quote from Doug Innocenti, CIO/CSO at MoonPay, on using Cato AI Security.Quote: "Having the right AI security solution gave me the confidence to say yes to the business."Source: [catonetworks.com](https://www.catonetworks.com/platform/ai-security-for-applications/), read 2026-11-16. - **aim-security-36** (vendor): Cato's 2026-08-03 launch release for Agentic Threat Prevention quotes Frank Dickson, group vice president at IDC, on AI-assisted attacks.Quote: "An entire attack can require less time than is required to investigate an alert"Source: [catonetworks.com](https://www.catonetworks.com/news/cato-launches-agentic-threat-prevention-to-predict-mitigate-ai-assisted-attacks/), read 2026-11-16. - **aim-security-37** (vendor): On 2026-10-06 Cato said it earned the AWS Security Competency in AI Security (Agentic AI category), one of fewer than 25 AWS partners to do so.Quote: "Cato is one of fewer than 25 AWS partners to earn the AI Security Competency"Source: [catonetworks.com](https://www.catonetworks.com/news/cato-achieves-aws-security-competency-in-ai-security/), read 2026-11-16. - **aim-security-38** (vendor): On 2026-10-27 Cato launched Cato AI Endpoint; the release quotes Alfred Sobrinho, CIO of PlayPower, and Chris DePuy of 650 Group.Quote: "“AI is creating real productivity gains across our business, but it also introduces new concerns about visibility, data protection, and governance,”"Source: [catonetworks.com](https://www.catonetworks.com/news/cato-launches-cato-ai-endpoint-secure-ai-use-beyond-browser/), read 2026-11-16. - **aim-security-40** (vendor): Cato says endpoint enforcement covers coding agents and browser AI apps, naming Claude Code, Cursor and Codex.Quote: "Covers coding agents on the endpoint and AI apps in the browser, with enforcement for Claude Code, Cursor, and Codex."Source: [catonetworks.com](https://www.catonetworks.com/solutions/ai-security/), read 2026-11-16. - **aim-security-41** (vendor): Cato says AI Endpoint blocks an unapproved tool call from a local agent on the endpoint before it runs, while the rest of the agent's work continues.Quote: "Cato AI Endpoint blocks that tool call on the endpoint before it runs, and the rest of the agent’s work carries on."Source: [catonetworks.com](https://www.catonetworks.com/platform/cato-ai-endpoint/), read 2026-11-16. - **aim-security-44** (vendor): Cato says it ingests and governs activity from Amazon Bedrock workloads through API integration without extra software in the Bedrock environment.Quote: "Through API integration, it ingests and governs activity from Bedrock-based workloads, extending visibility and policy-driven governance without requiring additional software in the Bedrock environment."Source: [catonetworks.com](https://www.catonetworks.com/news/cato-achieves-aws-security-competency-in-ai-security/), read 2026-11-16. - **aim-security-49** (vendor): Cato says AI Security for Applications deploys via proxy interception, AI-gateway integrations or APIs into model platforms and developer tools.Quote: "Deploy via proxy interception, AI-gateway integrations, or APIs into model platforms and developer tools."Source: [catonetworks.com](https://www.catonetworks.com/platform/ai-security-for-applications/), read 2026-11-16. - **aim-security-50** (vendor): Cato says applications are protected through the Cato cloud, the customer's existing AI gateway, or on-prem with Outposts.Quote: "Protects apps through the Cato cloud, the AI gateway you already run, or on-prem with Outposts."Source: [catonetworks.com](https://www.catonetworks.com/solutions/ai-security/), read 2026-11-16. - **aim-security-51** (vendor): Cato says it applies runtime controls to stop unauthorized or unsafe agent actions before execution.Quote: "Cato applies runtime controls to govern access and stop unauthorized or unsafe actions before execution."Source: [catonetworks.com](https://www.catonetworks.com/platform/agentic-ai-security/), read 2026-11-16. - **aim-security-52** (vendor): Cato says it inspects prompts, responses, tool calls and tool actions and blocks risky actions before they execute.Quote: "Block risky actions before they execute"Source: [catonetworks.com](https://www.catonetworks.com/platform/agentic-ai-security/), read 2026-11-16. - **aim-security-53** (third-party): In a PRNewswire release dated 2026-07-27 Cato says its engine inspects and enforces policy on tool inputs and outputs on all major agentic platforms.Quote: "The Cato AI engine now inspects and enforces policies on tool inputs and outputs on all major agentic platforms"Source: [prnewswire.com](https://www.prnewswire.com/news-releases/cato-networks-exceeds-415m-in-arr-fueled-by-42-yoy-growth-and-enterprise-demand-302833667.html), read 2026-11-16. - **aim-security-55** (vendor): Cato says one control plane enforces consistent policy across every app and agent with a single source of truth for compliance and incident response.Quote: "One policy across the entire AI surface"Source: [catonetworks.com](https://www.catonetworks.com/platform/ai-security-for-applications/), read 2026-11-16. - **aim-security-56** (vendor): Cato says AI and security events sit together in one data lake for audit, analytics and response.Quote: "AI and security events sit together for audit, analytics and response."Source: [catonetworks.com](https://www.catonetworks.com/solutions/ai-security/), read 2026-11-16. - **aim-security-57** (none-found): None found on 2026-11-16: the Cato AI security solution page does not name SIEM, SOAR, EDR vendor, identity provider or ticketing integrations for AI events.Source: [catonetworks.com](https://www.catonetworks.com/solutions/ai-security/), read 2026-11-16. - **aim-security-59** (vendor): Cato says the AI-Firewall detects jailbreaks, prompt injection and off-policy responses, and blocks or redacts them inline.Quote: "Detect jailbreaks, prompt injection, and off-policy responses"Source: [catonetworks.com](https://www.catonetworks.com/platform/ai-security-for-applications/), read 2026-11-16. - **aim-security-60** (vendor): Cato lists runtime control for applications as blocking prompt injection and jailbreaks.Quote: "Block prompt injection and jailbreaks"Source: [catonetworks.com](https://www.catonetworks.com/solutions/ai-security/), read 2026-11-16. - **aim-security-61** (vendor): Cato says it stops unsafe outputs before they reach users or external models.Quote: "Stop unsafe outputs before they reach users or external models"Source: [catonetworks.com](https://www.catonetworks.com/platform/ai-security-for-applications/), read 2026-11-16. - **aim-security-62** (none-found): None found on 2026-11-16: the Cato AI Endpoint and Agents pages name tool calls and prompts as blocked actions, but do not document blocking shell commands, file reads or writes, package installs, network calls or credential access as named action types.Source: [catonetworks.com](https://www.catonetworks.com/platform/cato-ai-endpoint/), read 2026-11-16. - **aim-security-63** (vendor): Cato says its AI Security Engine evaluates context across the full interaction to stop hidden risks before agent action.Quote: "Protect agents with the Cato AI Security Engine, which detects AI-native risk with industry-leading precision, recall, and low latency while evaluating context across the full interaction."Source: [catonetworks.com](https://www.catonetworks.com/platform/agentic-ai-security/), read 2026-11-16. - **aim-security-64** (vendor): Cato describes fast checks in milliseconds for every interaction, then deeper models only for cases needing a closer look; no figure is given.Quote: "Fast checks · milliseconds"Source: [catonetworks.com](https://www.catonetworks.com/solutions/ai-security/), read 2026-11-16. - **aim-security-65** (vendor): Cato claims high-fidelity, low-latency protection with minimal false positives, with no number published.Quote: "High-fidelity, low-latency protection with minimal false positives"Source: [catonetworks.com](https://www.catonetworks.com/platform/ai-security-for-applications/), read 2026-11-16. - **aim-security-66** (none-found): None found on 2026-11-16: no published latency figure (for example in milliseconds at a percentile) and no documented fail-open or fail-closed behavior or timeout on the Cato AI Security for Applications and AI Endpoint pages.Source: [catonetworks.com](https://www.catonetworks.com/platform/ai-security-for-applications/), read 2026-11-16. - **aim-security-67** (vendor): Cato says it monitors prompts, tool calls and actions and logs every interaction to assess risk.Quote: "Cato monitors agent behavior: prompts, tool calls, and actions: logging every interaction to assess risk."Source: [catonetworks.com](https://www.catonetworks.com/platform/agentic-ai-security/), read 2026-11-16. - **aim-security-68** (vendor): Cato says endpoint and network AI activity sit in one dashboard that ties interactions to user, endpoint and data, with blocked and monitored results.Quote: "Endpoint and network AI activity sit in one dashboard, and the same Cato AI engine inspects both."Source: [catonetworks.com](https://www.catonetworks.com/platform/cato-ai-endpoint/), read 2026-11-16. - **aim-security-69** (vendor): Cato lists logging violations for auditable evidence and full audit trails for investigation.Quote: "Preserve full audit trails for investigation"Source: [catonetworks.com](https://www.catonetworks.com/platform/ai-security-for-end-users/), read 2026-11-16. - **aim-security-70** (vendor): Cato says AI Security for Applications can run on-prem or in the customer cloud with AI Security Outposts so prompts stay in the customer environment.Quote: "Run AI Security for Applications on-prem or in your own cloud with AI Security Outposts. Prompts stay in your environment."Source: [catonetworks.com](https://www.catonetworks.com/solutions/ai-security/), read 2026-11-16. - **aim-security-71** (vendor): Cato says it embeds Aim's inspection into its SPACE enforcement layer to analyze prompts, responses, agent workflows and model outputs in real time.Quote: "we gain the ability to analyze AI interactions in real time: prompts, responses, agent workflows, and model outputs."Source: [catonetworks.com](https://www.catonetworks.com/blog/securing-ai-transformation-why-cato-acquired-aim-security/), read 2026-11-16. - **aim-security-72** (vendor): Dated news: on 2026-08-03 Cato launched Agentic Threat Prevention, an autonomous-agent capability to predict attack paths; it is network threat prevention, not AI runtime control.Quote: "today announced the launch of Cato Agentic Threat Prevention, a new capability that deploys autonomous agents to predict likely attack paths"Source: [catonetworks.com](https://www.catonetworks.com/news/cato-launches-agentic-threat-prevention-to-predict-mitigate-ai-assisted-attacks/), read 2026-11-16. - **aim-security-73** (vendor): Cato says it can discover agents and MCP servers, then block risky prompts, actions and tool calls.Quote: "Discover agents and MCP servers, then block risky prompts, actions, and tool calls."Source: [catonetworks.com](https://www.catonetworks.com/platform/ai-security-for-end-users/), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [Zenity profile](/vendors/zenity/) - [Bay profile](/vendors/bay/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # Bay AI runtime security URL: https://airuntimesecuritytools.com/vendors/bay/ Updated: 2026-11-16 Bay makes allow, ask, deny or steer decisions on the device itself. Bay says those decisions take under 4 ms and finish before the action completes, with no cloud round trip. Its stops cover credential access, risky shell commands, untrusted MCP connections, browser prompt data and prompt injection, and third-party press confirms the pre-execution blocking. Deployment is one script through an existing MDM or EDR in under five minutes, also confirmed by press. Decision logs export in OCSF to a SIEM. Gaps from the 16 Nov 2026 evidence: package install and network call blocking, fail-open or fail-closed behavior, and model or application layer enforcement are none found. Bay is a young company that left stealth in 2026, publishes no price and has little independent testing behind its claims. Bay is a young company that left stealth in 2026, publishes no price and has little independent testing behind its claims. ## Strength and weakness for AI Runtime Security Tools **Strength:** Decisions are made on the device in under 4ms, before the action completes (bay-49, bay-47). **Weakness:** None found on 2026-11-16 for fail-open or fail-closed behavior, or package-install and network-call blocking (bay-50, bay-57). ## Scores and evidence Bay is joint 4th in the overall ranking at 62.5 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Bay: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Action types stopped before they complete | 8 of 10 | 30 | The vendor claims stops for credential access, risky shell commands, untrusted MCP connections, browser prompt data and prompt injection, and third-party press confirms pre-execution blocking. Package install and network call blocking: none found. | [bay-48](#fact-bay-48), [bay-23](#fact-bay-23), [bay-51](#fact-bay-51), [bay-53](#fact-bay-53), [bay-55](#fact-bay-55), [bay-54](#fact-bay-54), [bay-57](#fact-bay-57) | | Layers covered | 6 of 10 | 20 | Endpoint and agent action enforcement is confirmed by third-party press, with browser prompt controls. Model or application layer enforcement for AI apps: none found. | [bay-46](#fact-bay-46), [bay-47](#fact-bay-47), [bay-23](#fact-bay-23), [bay-52](#fact-bay-52) | | Latency and failure mode | 5 of 10 | 15 | The vendor claims on-device decisions in under 4 ms. Fail-open or fail-closed behavior: none found. | [bay-49](#fact-bay-49), [bay-47](#fact-bay-47), [bay-50](#fact-bay-50) | | Deployment | 8 of 10 | 15 | One MDM or EDR script in under five minutes, confirmed by third-party press. Supported operating systems: none found. | [bay-16](#fact-bay-16), [bay-17](#fact-bay-17), [bay-19](#fact-bay-19), [bay-20](#fact-bay-20), [bay-21](#fact-bay-21) | | Detection and response | 7 of 10 | 5 | The vendor claims decision logs, OCSF SIEM export and session timelines. SOAR or ticketing integrations: none found. | [bay-61](#fact-bay-61), [bay-62](#fact-bay-62), [bay-63](#fact-bay-63), [bay-43](#fact-bay-43), [bay-65](#fact-bay-65) | | Transparency | 0 of 10 | 10 | Documentation is observed behind sign-in. Public documentation, a published price and published research: none found. | [bay-34](#fact-bay-34), [bay-35](#fact-bay-35), [bay-36](#fact-bay-36) | | Proof | 7 of 10 | 5 | Third-party press confirms funding and named customers, but notes unverified claims. Analyst coverage: none found. | [bay-06](#fact-bay-06), [bay-07](#fact-bay-07), [bay-08](#fact-bay-08), [bay-09](#fact-bay-09) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | joint 4th | 62.5 | | [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) | 3rd | 67.5 | | [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) | joint 3rd | 63.5 | | [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) | joint 4th | 67.0 | | [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) | joint 3rd | 66.5 | | [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) | 2nd | 69.0 | ## Product, deployment, price and owner - Product assessed: Bay Agent Runtime Enforcement (Bay Enforcer, Lighthouse policy engine) - Deployment: Agentless: one script through an existing MDM or EDR runs Wave, an ephemeral binary; Bay Enforcer, a static binary, decides on the device via native hook; no proxy (vendor claims) - Time to deploy: Under 5 minutes, as claimed by the vendor; proof of value in under 14 days (SecurityWeek quote) - Price: none found on 2026-11-16 - Owner: Bay AI, Inc. - Website: [bay.io](https://bay.io/?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor) ## Facts and sources - **bay-01** (vendor): Bay AI, Inc. publishes the Bay site and states its copyright on every page.Quote: "© 2026 Bay AI, Inc. All rights reserved."Source: [bay.io](https://www.bay.io/about?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-02** (vendor): Bay says it was founded by three engineers from endpoint and API security.Quote: "Bay was founded by three engineers who spent their careers solving the hardest problems in endpoint and API security."Source: [bay.io](https://www.bay.io/about?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-03** (third-party): On 2026-11-10 Bay named its leadership: Yuval Zacharia is CEO and co-founder, Aner Morag is CPO and Boris Danilovich is VP of R&D, per SiliconANGLE.Quote: "Zacharia founded Bay with Chief Product Officer Aner Morag and Vice President of Research and Development Boris Danilovich."Source: [siliconangle.com](https://siliconangle.com/2026/11/10/ai-security-startup-bay-launches-with-50m-to-vet-agent-actions-before-they-run/), read 2026-11-16. - **bay-04** (vendor): Bay says it was founded in 2026 and is headquartered in Tel Aviv; its press release is dated 2026-11-10.Quote: "Founded in 2026, Bay is headquartered in Tel Aviv and serves customers in the United States and Europe."Source: [bay.io](https://www.bay.io/blog/bay-emerges-from-stealth-with-50m-after-fortune-100-companies-used-its-platform-to-triple-autonomous-ai-agent-use-2?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-05** (vendor): On 2026-11-10 Bay said it had raised $50 million in total funding.Quote: "Bay has raised $50 million in total funding."Source: [bay.io](https://www.bay.io/blog/bay-emerges-from-stealth-with-50m-after-fortune-100-companies-used-its-platform-to-triple-autonomous-ai-agent-use-2?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-06** (third-party): On 2026-11-11 SecurityWeek reported Bay emerged from stealth with a $37 million Series A led by Battery Ventures plus a $13 million seed.Quote: "emerged from stealth with $50 million funding (comprising a $37 million Series A round led by Battery Ventures"Source: [securityweek.com](https://www.securityweek.com/bay-raises-50-million-to-control-what-enterprise-ai-agents-can-do/), read 2026-11-16. - **bay-07** (third-party): On 2026-11-10 SiliconANGLE reported Micron, Wix, ClickUp and Peloton are among dozens of Bay enterprise customers.Quote: "Micron Technology Inc., Wix.com Ltd., ClickUp Inc. and Peloton Interactive Inc. are among the dozens of enterprise customers signed in the nine months since Bay was founded."Source: [siliconangle.com](https://siliconangle.com/2026/11/10/ai-security-startup-bay-launches-with-50m-to-vet-agent-actions-before-they-run/), read 2026-11-16. - **bay-08** (third-party): The Next Web noted that Bay did not name its Fortune 100 customers or say how it measured the tripling claim, and that one named customer is also an investor.Quote: "The company did not name those customers or say how it measured the figure, and TNW has not verified it."Source: [thenextweb.com](https://thenextweb.com/news/bay-50m-ai-agent-security-stealth-battery-ventures), read 2026-11-16. - **bay-09** (none-found): None found on 2026-11-16: independent analyst coverage of Bay (Gartner, Forrester or similar) on the Bay home page or blog.Source: [bay.io](https://bay.io/?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-10** (vendor): Bay says it is SOC 2 and ISO 27001 certified.Quote: "SOC 2 & ISO 27001 Certified"Source: [bay.io](https://www.bay.io/about?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-16** (vendor): Bay says it deploys with one script through an existing MDM, with no proxy and nothing installed by users.Quote: "Bay deploys with one script through an existing MDM. It does not require a proxy to route traffic through and does not require anything to be installed by users."Source: [bay.io](https://www.bay.io/platform?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-17** (vendor): Bay says it is agentless and deploys via MDM or EDR using Wave, an ephemeral binary, with no persistent agent and no proxy.Quote: "Bay is agentless: it deploys through your existing MDM or EDR using Wave, an ephemeral binary, so there’s no persistent agent to maintain and no proxy in the path."Source: [bay.io](https://www.bay.io/platform/ai-endpoint-security?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-18** (vendor): Bay says the endpoint inventory appears in under 5 minutes.Quote: "Bay brings it all into view in under 5 minutes and keeps it in bounds, without slowing anyone down."Source: [bay.io](https://www.bay.io/platform/ai-endpoint-security?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-19** (third-party): On 2026-11-10 SiliconANGLE reported Bay installs through the customer's existing MDM or EDR and the company says it takes under five minutes.Quote: "Bay installs through the mobile device management or endpoint detection and response tools a customer already has, and the company says it takes under five minutes."Source: [siliconangle.com](https://siliconangle.com/2026/11/10/ai-security-startup-bay-launches-with-50m-to-vet-agent-actions-before-they-run/), read 2026-11-16. - **bay-20** (third-party): On 2026-11-11 SecurityWeek quoted Bay saying a fleet deploys in under five minutes and a proof of value completes in under 14 days.Quote: "complete a proof of value in less than 14 days."Source: [securityweek.com](https://www.securityweek.com/bay-raises-50-million-to-control-what-enterprise-ai-agents-can-do/), read 2026-11-16. - **bay-21** (none-found): None found on 2026-11-16: a list of supported operating systems for the Bay endpoint scanner, on the endpoint product page and the platform page.Source: [bay.io](https://www.bay.io/platform/ai-endpoint-security?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-23** (vendor): Bay says browser guardrails keep personal data, secrets and source code out of prompts, with allow, warn, steer or block per team.Quote: "Keep personal data, secrets and source code out of prompts, and decide what to, Allow, Warn, Steer or Block, tailored to each team."Source: [bay.io](https://www.bay.io/platform/browser?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-24** (vendor): Bay says it can allow or block specific AI domains and extensions and stop risky agentic actions inside the browser.Quote: "Bay lets you allow or block specific AI domains and extensions, and stops risky agentic actions inside the browser"Source: [bay.io](https://www.bay.io/platform/browser?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-34** (observed): Bay documentation at docs.bay.io redirects to a sign-in page and is open only to signed-in Bay users.Quote: "The Bay documentation is open only to signed-in Bay users."Source: [docs.bay.io](https://docs.bay.io/?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-35** (none-found): None found on 2026-11-16: a published price. The site has no pricing page (bay.io/pricing returns 404) and the demo page only offers a 30-minute demo.Source: [bay.io](https://www.bay.io/demo?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-36** (none-found): None found on 2026-11-16: published research, threat reports or open-source code; the Bay blog lists only two announcement posts dated 2026-11-10.Source: [bay.io](https://www.bay.io/blog?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-41** (vendor): Bay says it shows which credentials, secrets and permissions each cloud agent can reach so risky ones can be locked down.Quote: "See which credentials, secrets and permissions each cloud agent can reach, and lock down the risky ones before a long-running task puts them to use."Source: [bay.io](https://www.bay.io/platform/cloud-workspaces?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-43** (vendor): Bay says it attributes each action to the person or the agent acting for them and records the agent's permission mode.Quote: "Bay attributes each action to the person who took it or the agent that acted for them, and records the permission mode the agent was running in."Source: [bay.io](https://www.bay.io/ai-agent-investigation?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-44** (vendor): Bay says it uses no proxy; decisions are made on the device.Quote: "No proxy: every decision is made on the device."Source: [bay.io](https://www.bay.io/agent-runtime-enforcement?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-46** (third-party): On 2026-11-10 SiliconANGLE reported the AI tool hands each proposed action to Bay Enforcer, a static binary, through a native hook on the machine.Quote: "Each time an agent proposes an action, the AI tool running it hands the request to Bay Enforcer, a static binary, through a native hook."Source: [siliconangle.com](https://siliconangle.com/2026/11/10/ai-security-startup-bay-launches-with-50m-to-vet-agent-actions-before-they-run/), read 2026-11-16. - **bay-47** (vendor): Bay says runtime enforcement decides allow, ask, deny or steer on every action, locally and in milliseconds with no cloud round-trip.Quote: "Bay reads each agent’s whole session to understand its intent, then decides on every action: allow, ask, deny or steer."Source: [bay.io](https://www.bay.io/agent-runtime-enforcement?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-48** (vendor): Bay says it acts on credential access, risky shell commands and untrusted MCP server connections in milliseconds.Quote: "When an agent reaches for credentials, runs a risky shell command or connects to an untrusted MCP server, Bay allows, asks, denies or steers the action in milliseconds"Source: [bay.io](https://www.bay.io/platform/ai-endpoint-security?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-49** (vendor): Bay says endpoint decisions are made on the device in under 4ms.Quote: "allowing, asking, denying or steering it right on the device in under 4ms"Source: [bay.io](https://www.bay.io/platform/ai-endpoint-security?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-50** (none-found): None found on 2026-11-16: published behavior when the Bay Enforcer fails or times out (fail open or fail closed), on the runtime enforcement page.Source: [bay.io](https://www.bay.io/agent-runtime-enforcement?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-51** (vendor): Bay says it can prevent prompt injection attempts at runtime.Quote: "Bay can also prevent prompt injection attempts at runtime"Source: [bay.io](https://www.bay.io/platform?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-52** (third-party): On 2026-11-10 SiliconANGLE reported Bay argues first-generation tools that filter prompts and responses miss where damage happens; Bay focuses on agent actions.Quote: "First-generation AI security tools mostly filter prompts and responses."Source: [siliconangle.com](https://siliconangle.com/2026/11/10/ai-security-startup-bay-launches-with-50m-to-vet-agent-actions-before-they-run/), read 2026-11-16. - **bay-53** (third-party): On 2026-11-10 SiliconANGLE reported that anything Bay judges too risky never runs.Quote: "Anything judged too risky never runs."Source: [siliconangle.com](https://siliconangle.com/2026/11/10/ai-security-startup-bay-launches-with-50m-to-vet-agent-actions-before-they-run/), read 2026-11-16. - **bay-54** (third-party): On 2026-11-11 SecurityWeek reported Bay decides in milliseconds to allow, refer to the user, redirect or deny each action.Quote: "decides in milliseconds whether to allow the action, refer it to the user, redirect it based on policy, or simply deny it"Source: [securityweek.com](https://www.securityweek.com/bay-raises-50-million-to-control-what-enterprise-ai-agents-can-do/), read 2026-11-16. - **bay-55** (third-party): On 2026-11-10 The Next Web reported Bay checks every action an agent proposes before it runs, then allows, blocks or asks a person.Quote: "It checks every action an agent proposes before it runs, then allows it, blocks it or asks a person."Source: [thenextweb.com](https://thenextweb.com/news/bay-50m-ai-agent-security-stealth-battery-ventures), read 2026-11-16. - **bay-56** (vendor): Bay says it steps in only when credentials are touched and leaves normal work alone.Quote: "It steps in only when credentials are touched and leaves normal work alone."Source: [bay.io](https://www.bay.io/platform?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-57** (none-found): None found on 2026-11-16: named blocking of package installs or network calls, or inspection of tool-call arguments, on the endpoint and runtime pages.Source: [bay.io](https://www.bay.io/platform/ai-endpoint-security?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-58** (vendor): Bay says it applies intent-aware enforcement on every tool call in cloud agents.Quote: "Intent-aware enforcement on every tool call."Source: [bay.io](https://www.bay.io/platform/cloud-workspaces?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-59** (vendor): Bay says any policy can run in simulation to show what it would allow, ask, deny or steer before users are affected.Quote: "Run any policy in simulation to see what each rule would allow, ask, deny or steer, before a single user is affected."Source: [bay.io](https://www.bay.io/agent-runtime-enforcement?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-61** (vendor): Bay says every decision is logged against the policy behind it for auditors.Quote: "Every decision is logged against the policy behind it, ready for your auditors."Source: [bay.io](https://www.bay.io/platform?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-62** (vendor): Bay says events reach the customer SIEM in OCSF format.Quote: "Events arrive in your SIEM in OCSF format, with enough context that analysts can triage without opening another console."Source: [bay.io](https://www.bay.io/ai-agent-investigation?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-63** (vendor): Bay says it records every prompt, tool call and system action on one timeline per session.Quote: "Every prompt, tool call and system action on one timeline, so you can see how an agent got from its first instruction to its last action."Source: [bay.io](https://www.bay.io/ai-agent-investigation?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. - **bay-65** (none-found): None found on 2026-11-16: named EDR, SOAR or ticketing integrations beyond SIEM export, on the investigation page.Source: [bay.io](https://www.bay.io/ai-agent-investigation?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=vendor), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [Aim Security profile](/vendors/aim-security/) - [Check Point profile](/vendors/check-point/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # Check Point AI runtime security URL: https://airuntimesecuritytools.com/vendors/check-point/ Updated: 2026-11-16 Check Point screens prompts, outputs and agent tool calls in the request path, and publishes a latency claim of under 50 ms. Its guardrails run as managed SaaS or self-hosted on Kubernetes, Docker or air-gapped, integrated through a Guard API, and its workforce product adds a browser extension for data paste and upload. Documentation and an OpenAPI spec are public without sign-in. The calling application carries out the block, so enforcement depends on how it is integrated. Gaps from 16 Nov 2026: blocking of shell commands, package installs and network calls on the machine, SIEM or EDR integrations by name and a fail-open or fail-closed behavior are none found. ## Strength and weakness for AI Runtime Security Tools **Strength:** Runtime screening of prompts, outputs and agent tool calls with a published sub-50 ms latency claim and SaaS, self-hosted or air-gapped deployment (check-point-47, check-point-31, check-point-26). **Weakness:** None found on 2026-11-16 for machine-level blocking of shell commands, package installs or network calls, nor a published fail-open or fail-closed behavior (check-point-44, check-point-49). ## Scores and evidence Check Point is 9th in the overall ranking at 55.0 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Check Point: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Action types stopped before they complete | 5 of 10 | 30 | The vendor says it stops data paste and upload, unsafe tool calls, prompt injection and harmful output, but the calling application carries out the block. Blocking of shell, file, package, network and credential actions on the machine: none found. | [check-point-13](#fact-check-point-13), [check-point-25](#fact-check-point-25), [check-point-42](#fact-check-point-42), [check-point-43](#fact-check-point-43), [check-point-50](#fact-check-point-50), [check-point-44](#fact-check-point-44) | | Layers covered | 6 of 10 | 20 | Vendor documentation covers the model, application and agent tool-call layers. Enforcement of actions on the machine: none found. | [check-point-22](#fact-check-point-22), [check-point-23](#fact-check-point-23), [check-point-45](#fact-check-point-45), [check-point-08](#fact-check-point-08), [check-point-44](#fact-check-point-44) | | Latency and failure mode | 5 of 10 | 15 | The vendor claims under 50 ms at runtime. Published fail-open or fail-closed behavior and timeouts: none found. | [check-point-47](#fact-check-point-47), [check-point-48](#fact-check-point-48), [check-point-49](#fact-check-point-49) | | Deployment | 6 of 10 | 15 | Managed SaaS or self-hosted on Kubernetes, Docker or air-gapped, plus a browser extension. A time to deploy is published only for the workforce product; none found for agent and API protection. | [check-point-30](#fact-check-point-30), [check-point-31](#fact-check-point-31), [check-point-12](#fact-check-point-12), [check-point-32](#fact-check-point-32) | | Detection and response | 4 of 10 | 5 | The vendor claims unified audit trails, a logs API and a policy simulator. Named SIEM, EDR or ticketing integrations: none found. | [check-point-33](#fact-check-point-33), [check-point-36](#fact-check-point-36), [check-point-34](#fact-check-point-34), [check-point-35](#fact-check-point-35) | | Transparency | 6 of 10 | 10 | Documentation, OpenAPI spec and research are public without sign-in. Published price: none found. | [check-point-37](#fact-check-point-37), [check-point-39](#fact-check-point-39), [check-point-38](#fact-check-point-38) | | Proof | 7 of 10 | 5 | Third-party press covers the acquisition and platform launch, and the vendor names customers. Independent analyst coverage: none found. | [check-point-02](#fact-check-point-02), [check-point-03](#fact-check-point-03), [check-point-40](#fact-check-point-40), [check-point-41](#fact-check-point-41) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 9th | 55.0 | | [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) | 12th | 54.0 | | [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) | 8th | 54.5 | | [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) | 12th | 55.0 | | [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) | 12th | 51.0 | | [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) | 9th | 56.0 | ## Product, deployment, price and owner - Product assessed: Check Point AI Guardrails (Guard API) and AI Agent Security runtime protection - Deployment: Managed SaaS or self-hosted (Kubernetes Helm, Docker, air-gapped) for AI Guardrails via API; browser extension plus optional desktop agent for Workforce AI Security; agent platform connectors for discovery - Time to deploy: Workforce AI Security says it deploys instantly through a browser extension; none found on 2026-11-16 for AI Agent Security and AI Guardrails - Price: none found on 2026-11-16 - Owner: Check Point Software Technologies Ltd. - Website: checkpoint.com ## Facts and sources - **check-point-01** (vendor): Check Point Software Technologies announced on 2025-09-16 that it agreed to acquire Lakera, an AI-native security platform; the press release says Lakera products became Check Point AI Red Teaming and Check Point AI Agent Security.Quote: "Its flagship solutions, Check Point AI Red Teaming and Check Point AI Agent Security, provide pre-deployment posture assessments and real-time runtime enforcement"Source: [checkpoint.com](https://www.checkpoint.com/press-releases/check-point-acquires-lakera-to-deliver-end-to-end-ai-security-for-enterprises/), read 2026-11-16. - **check-point-02** (third-party): SecurityWeek reported on 2025-09-16 that Check Point planned to acquire Lakera, and that on closing Lakera would form the basis of a Global Center of Excellence for AI Security.Quote: "Upon closing, Lakera will form the basis of Check Point’s new Global Center of Excellence for AI Security"Source: [securityweek.com](https://www.securityweek.com/check-point-to-acquire-ai-security-firm-lakera/), read 2026-11-16. - **check-point-03** (third-party): Help Net Security reported on 2026-03-24 that Check Point launched the AI Defense Plane, built on the Lakera and Cyata acquisitions, with workforce, AI application and agent, and red teaming modules.Quote: "strengthened by technologies from ThreatCloud AI and the company’s recent Lakera and Cyata acquisitions"Source: [helpnetsecurity.com](https://www.helpnetsecurity.com/2026/03/24/check-point-ai-defense-plane/), read 2026-11-16. - **check-point-04** (vendor): Check Point lists Workforce AI Security, AI Red Teaming and AI Agent Security as its AI Security products, and describes an AI Defense Plane to discover, protect and govern AI.Quote: "Check Point delivers one platform to discover, protect, and govern every AI interaction from employees to agents."Source: [checkpoint.com](https://www.checkpoint.com/ai-security/), read 2026-11-16. - **check-point-05** (vendor): The Lakera documentation site is now branded Check Point AI Security, with products AI Agent Security, AI Guardrails and AI Red Teaming.Quote: "AI Guardrails is the runtime layer of AI Agent Security, and is also sold on its own"Source: [docs.lakera.ai](https://docs.lakera.ai/docs/agent-security), read 2026-11-16. - **check-point-06** (vendor): The Lakera website still operates as lakera.ai and describes workforce AI security, AI agent security and AI red teaming products.Quote: "Shadow AI discovery across apps and browsers"Source: [lakera.ai](https://www.lakera.ai), read 2026-11-16. - **check-point-07** (vendor): A 2026-10-01 Check Point press release names Adam Ely as General Manager, AI Security at Check Point.Quote: "said Adam Ely, General Manager, AI Security at Check Point Software Technologies"Source: [checkpoint.com](https://www.checkpoint.com/press-releases/clalit-health-services-and-check-point-announce-strategic-partnership-to-secure-ai-adoption-across-healthcare-environments/), read 2026-11-16. - **check-point-08** (vendor): Check Point says Workforce AI Security covers desktop agents, code assistants such as GitHub Copilot and Cursor, and MCP-connected workflows; the Enterprise tier covers web, desktop, MSP and IDEs.Quote: "The Enterprise tier covers web, desktop, MSP, and IDEs, while the Essentials tier focuses on web applications."Source: [checkpoint.com](https://www.checkpoint.com/ai-security/workforce-ai-security/), read 2026-11-16. - **check-point-10** (vendor): Check Point says endpoint coverage comes from an optional desktop agent, while the base deployment is a browser extension.Quote: "Desktop agent deployment is also available for broader coverage across device-level AI usage."Source: [checkpoint.com](https://www.checkpoint.com/ai-security/workforce-ai-security/), read 2026-11-16. - **check-point-12** (vendor): Check Point says Workforce AI Security deploys through a lightweight browser extension with no architecture changes.Quote: "Workforce AI Security deploys instantly through a lightweight browser extension, no complex setup, no downtime, and no architecture changes required."Source: [checkpoint.com](https://www.checkpoint.com/ai-security/workforce-ai-security/), read 2026-11-16. - **check-point-13** (vendor): Check Point says Workforce AI Security protects data in prompts with inline redaction, copy and paste restrictions, file and image redaction with OCR, and user guidance.Quote: "Protections include inline redaction, copy/paste restrictions, file and image redaction with OCR, and interactive user guidance."Source: [checkpoint.com](https://www.checkpoint.com/ai-security/workforce-ai-security/), read 2026-11-16. - **check-point-22** (vendor): Check Point documents the Guard API as the single integration point that screens prompts, model outputs, tool calls, tool responses and tool descriptions for an agent workflow.Quote: "applied across the agentic workflow via the Guard API: user prompts, model outputs, tool calls, tool responses, and tool descriptions"Source: [docs.lakera.ai](https://docs.lakera.ai/docs/agent-security), read 2026-11-16. - **check-point-23** (vendor): Check Point says AI Agent Security acts as an inline enforcement layer that inspects prompts, responses, external content and agent tool calls before risky actions run.Quote: "AI Agent Security acts as an inline enforcement layer for AI agent interactions."Source: [checkpoint.com](https://www.checkpoint.com/ai-security/ai-agent-security/), read 2026-11-16. - **check-point-25** (vendor): Check Point says AI Agent Security evaluates agent tool calls in context to help block unsafe or unauthorized behavior at runtime.Quote: "Evaluate agent tool calls and actions in context, helping block unsafe or unauthorized behavior at runtime."Source: [checkpoint.com](https://www.checkpoint.com/ai-security/ai-agent-security/), read 2026-11-16. - **check-point-26** (vendor): Check Point docs say the Dangerous Deviation detector (beta) flags agent tool calls that cause data leakage, privilege escalation or system destruction with no warrant in the trusted conversation.Quote: "Access and privilege escalation : public sharing of resources, sudo or admin grants, widened ACLs, or attempts to escalate permissions beyond a sandbox."Source: [docs.lakera.ai](https://docs.lakera.ai/docs/agent-behavior-defense), read 2026-11-16. - **check-point-28** (vendor): A Check Point research blog dated 2026-09-15 describes a sidecar control monitor that prevents harmful agent actions before they execute at sub-100ms latency; the post is research, and its product status is not stated there.Quote: "prevents harmful actions before they execute in real time at sub-100ms latency"Source: [blog.checkpoint.com](https://blog.checkpoint.com/ai-security/synchronous-control-monitoring-preventing-harmful-agent-actions-in-real-time/), read 2026-11-16. - **check-point-30** (vendor): Check Point documents AI Guardrails as managed SaaS or self-hosted on Kubernetes (Helm), Docker, or air-gapped.Quote: "AI Guardrails is available as a fully managed SaaS solution or as a self-hosted deployment ."Source: [docs.lakera.ai](https://docs.lakera.ai/docs/selfhosting), read 2026-11-16. - **check-point-31** (vendor): Check Point documents that runtime protection integrates through the Guard API, with discovery and risk assessment managed in the portal.Quote: "Discovery and risk assessment are posture capabilities, managed in the portal."Source: [docs.lakera.ai](https://docs.lakera.ai/docs/agent-security), read 2026-11-16. - **check-point-32** (none-found): None found on 2026-11-16: looked for a published time to deploy for AI Agent Security and AI Guardrails in the docs quickstart and product pages; only Workforce AI claims "instantly".Source: [docs.lakera.ai](https://docs.lakera.ai/docs/api/guard), read 2026-11-16. - **check-point-33** (vendor): Check Point says the AI Defense Plane applies consistent policy with unified audit trails and compliance reporting.Quote: "Apply consistent policy across all AI interactions, with unified audit trails and compliance reporting."Source: [checkpoint.com](https://www.checkpoint.com/ai-security/), read 2026-11-16. - **check-point-34** (vendor): Check Point added a Policy Impact Simulator on 2026-04-10 that shows how sensitivity levels would have affected historical traffic.Quote: "an interactive tool that shows how different sensitivity levels and guardrail configurations would have affected your historical traffic"Source: [docs.lakera.ai](https://docs.lakera.ai/changelog), read 2026-11-16. - **check-point-35** (none-found): None found on 2026-11-16: looked for named SIEM, EDR, identity provider and ticketing integrations for AI Agent Security and AI Guardrails on the product pages, integration guide and docs index.Source: [docs.lakera.ai](https://docs.lakera.ai/docs/integration), read 2026-11-16. - **check-point-36** (vendor): Check Point docs provide a request latency percentiles chart on the Analytics page (2026-10-26) and a Guard logs API.Quote: "Request latency percentiles chart on the Analytics page"Source: [docs.lakera.ai](https://docs.lakera.ai/changelog), read 2026-11-16. - **check-point-37** (vendor): Check Point publishes AI Security documentation, an OpenAPI spec and a changelog that are readable without sign-in.Quote: "Subscribe via RSS"Source: [docs.lakera.ai](https://docs.lakera.ai/changelog), read 2026-11-16. - **check-point-38** (none-found): None found on 2026-11-16: looked for a published price for AI Agent Security, Workforce AI Security and AI Guardrails on the Check Point AI Security pages and the Lakera site; the pages only offer demos.Source: [checkpoint.com](https://www.checkpoint.com/ai-security/), read 2026-11-16. - **check-point-39** (vendor): Check Point publishes AI agent security research on its blog, including a post dated 2026-09-15 on synchronous control monitoring.Quote: "Control monitors ingest the trace as context."Source: [blog.checkpoint.com](https://blog.checkpoint.com/ai-security/synchronous-control-monitoring-preventing-harmful-agent-actions-in-real-time/), read 2026-11-16. - **check-point-40** (vendor): The AI Agent Security page quotes a Dropbox security engineer and the Nubank CISO as customers of the Lakera technology.Quote: "The Lakera team has accelerated our GenAI journey, allowing us to create secure GenAI experiences at scale."Source: [checkpoint.com](https://www.checkpoint.com/ai-security/ai-agent-security/), read 2026-11-16. - **check-point-41** (vendor): On 2026-10-01 Check Point announced Clalit Health Services will deploy Check Point AI Security for real-time visibility and control over AI usage and agents.Quote: "Check Point AI Security will give Clalit real-time visibility and control over AI usage and AI agents across the organization."Source: [checkpoint.com](https://www.checkpoint.com/press-releases/clalit-health-services-and-check-point-announce-strategic-partnership-to-secure-ai-adoption-across-healthcare-environments/), read 2026-11-16. - **check-point-42** (vendor): Check Point says AI Application Security blocks prompt injection, jailbreaks and harmful outputs at runtime.Quote: "Block prompt injection, jailbreaks, and harmful outputs at runtime with low-latency enforcement."Source: [checkpoint.com](https://www.checkpoint.com/ai-security/), read 2026-11-16. - **check-point-43** (vendor): Check Point docs say AI Guardrails includes Prompt Defense, Content Moderation, Data Leakage Prevention, Malicious Links and Agent Behavior Defense.Quote: "Prompt Defense , Content Moderation , Data Leakage Prevention , Malicious Links , and Agent Behavior Defense"Source: [docs.lakera.ai](https://docs.lakera.ai/docs/agent-security), read 2026-11-16. - **check-point-44** (none-found): None found on 2026-11-16: looked for blocking of shell commands, file reads or writes, package installs, network calls and credential access on the machine by AI Guardrails or AI Agent Security in the docs index and product pages.Source: [docs.lakera.ai](https://docs.lakera.ai/docs/agent-behavior-defense), read 2026-11-16. - **check-point-45** (vendor): Check Point says its AI security covers workforce usage, applications at the model input and output edge, and agent tool calls and file access.Quote: "Protect model inputs and outputs at the application edge."Source: [checkpoint.com](https://www.checkpoint.com/ai-security/), read 2026-11-16. - **check-point-46** (vendor): Check Point says agents are covered by controls on tool calls, file access and autonomous runtime behavior.Quote: "Control tool calls, file access, and autonomous runtime behavior."Source: [checkpoint.com](https://www.checkpoint.com/ai-security/), read 2026-11-16. - **check-point-47** (vendor): Check Point claims under 50 ms latency at runtime.Quote: "<50 ms latency guaranteed at runtime"Source: [checkpoint.com](https://www.checkpoint.com/ai-security/), read 2026-11-16. - **check-point-48** (vendor): Check Point docs say AI Guardrails keeps requests under a latency cap regardless of policy or content length, without stating the cap value.Quote: "keep requests under a latency “cap”, regardless of policy or content length"Source: [docs.lakera.ai](https://docs.lakera.ai/docs/api/guard), read 2026-11-16. - **check-point-49** (none-found): None found on 2026-11-16: looked for a published fail-open or fail-closed behavior and timeout values for the Guard API in the Guard API page and integration guide.Source: [docs.lakera.ai](https://docs.lakera.ai/docs/api/guard), read 2026-11-16. - **check-point-50** (vendor): Check Point docs say the application decides the action on a flagged response, so blocking is implemented by the calling application.Quote: "your application determines the appropriate action"Source: [docs.lakera.ai](https://docs.lakera.ai/docs/integration), read 2026-11-16. - **check-point-51** (vendor): Check Point says the Guard API returns a boolean flagged result with an optional per-detector breakdown including confidence levels.Quote: "Optionally, a breakdown of the flagging decision can be returned in the response by setting "breakdown": true in the request"Source: [docs.lakera.ai](https://docs.lakera.ai/docs/api/guard), read 2026-11-16. - **check-point-52** (vendor): On 2026-07-30 Check Point announced a new AI Network Firewall, listed in its newsroom.The exact quote is on the source page.Source: [checkpoint.com](https://www.checkpoint.com/press-releases/), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [Bay profile](/vendors/bay/) - [Cisco profile](/vendors/cisco/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # Cisco AI runtime security URL: https://airuntimesecuritytools.com/vendors/cisco/ Updated: 2026-11-16 Cisco enforces at four points: network, gateway, SDK and endpoint hooks. AI Defense inspects traffic in the network and through a Kong gateway plugin, and DefenseClaw checks a prompt, shell command or tool call before the agent acts, answering allow, alert, ask or block. The Kong plugin fails closed by default, and an operator can set it to fail open, which makes Cisco one of few vendors with a written failure mode. Every decision is recorded and exported to OpenTelemetry destinations. Gaps from 16 Nov 2026: a published inline latency figure and an inline package install block are none found, and paid list pricing is not published. A free Explorer Edition covers red teaming. ## Strength and weakness for AI Runtime Security Tools **Strength:** Inline inspection of prompts, responses, tool calls and tool results through network, gateway, SDK and endpoint hooks with documented fail-closed defaults (cisco-02, cisco-05, cisco-07, cisco-10). **Weakness:** None found on 2026-11-16 for a published inline latency figure or an inline package-install block (cisco-08, cisco-52). ## Scores and evidence Cisco is 2nd in the overall ranking at 66.5 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Cisco: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Action types stopped before they complete | 7 of 10 | 30 | The vendor documents pre-execution checks on shell commands, tool calls and prompts, plus network destination, credential placeholder, DLP and harmful response controls. Package install blocking: none found, and it states a limit on file writes by child programs. | [cisco-10](#fact-cisco-10), [cisco-14](#fact-cisco-14), [cisco-15](#fact-cisco-15), [cisco-49](#fact-cisco-49), [cisco-51](#fact-cisco-51), [cisco-02](#fact-cisco-02), [cisco-20](#fact-cisco-20), [cisco-52](#fact-cisco-52) | | Layers covered | 8 of 10 | 20 | Network, gateway, SDK and endpoint hook enforcement is documented, and framework plugins are observed on GitHub. | [cisco-02](#fact-cisco-02), [cisco-05](#fact-cisco-05), [cisco-12](#fact-cisco-12), [cisco-11](#fact-cisco-11), [cisco-45](#fact-cisco-45) | | Latency and failure mode | 5 of 10 | 15 | Timeouts and fail-closed defaults are published. Inline latency figure: none found. | [cisco-06](#fact-cisco-06), [cisco-07](#fact-cisco-07), [cisco-09](#fact-cisco-09), [cisco-08](#fact-cisco-08) | | Deployment | 7 of 10 | 15 | Network enforcement, MDM packages, a Kong plugin and SDKs are documented, and integrations are observed on GitHub. The time to deploy comes only from a reported vendor claim. | [cisco-04](#fact-cisco-04), [cisco-38](#fact-cisco-38), [cisco-11](#fact-cisco-11), [cisco-44](#fact-cisco-44), [cisco-45](#fact-cisco-45) | | Detection and response | 8 of 10 | 5 | Every decision is recorded and exported to OpenTelemetry and named destinations, and third-party press confirms a telemetry connector. | [cisco-17](#fact-cisco-17), [cisco-18](#fact-cisco-18), [cisco-35](#fact-cisco-35), [cisco-53](#fact-cisco-53), [cisco-37](#fact-cisco-37) | | Transparency | 5 of 10 | 10 | Public docs and Apache-2.0 code, and third-party press covers a free red-teaming edition. Paid list price for runtime protection and published research: none found. | [cisco-29](#fact-cisco-29), [cisco-57](#fact-cisco-57), [cisco-56](#fact-cisco-56) | | Proof | 5 of 10 | 5 | A public company with one named customer voice. Analyst coverage: none found. | [cisco-58](#fact-cisco-58), [cisco-59](#fact-cisco-59), [cisco-54](#fact-cisco-54) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 2nd | 66.5 | | [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) | 2nd | 69.0 | | [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) | 2nd | 65.5 | | [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) | 2nd | 71.0 | | [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) | 2nd | 70.5 | | [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) | 3rd | 68.0 | ## Product, deployment, price and owner - Product assessed: Cisco AI Defense runtime protection, Python SDK and Kong plugin inspection, and DefenseClaw endpoint hooks - Deployment: Network-level enforcement without agents or libraries for AI Defense; Kong plugin, Python SDK, command-line MCP Scanner, and DefenseClaw hooks installed per user or by MDM package - Time to deploy: Cisco says DefenseClaw installs in about five minutes (SiliconANGLE); none found on 2026-11-16 for AI Defense - Price: none found on 2026-11-16 (paid AI Defense); Explorer Edition is described as free - Owner: Cisco Systems, Inc. - Website: [cisco.com](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html) ## Facts and sources - **cisco-01** (vendor): Cisco Systems, Inc. is the owner and copyright holder of DefenseClaw, the open-source agent governance layer published in the cisco-ai-defense GitHub organization alongside AI Defense.The exact quote is on the source page.Source: [github.com](https://github.com/cisco-ai-defense/defenseclaw), read 2026-11-16. - **cisco-02** (vendor): Cisco says AI Defense runtime protection puts guardrails in the network and blocks adversarial attacks and harmful responses in real time.Quote: "Protect production AI applications with guardrails embedded in the network. Block adversarial attacks and harmful responses in real time."Source: [cisco.com](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html), read 2026-11-16. - **cisco-03** (vendor): Cisco says AI Defense protects AI applications against prompt injection, denial of service and data leakage.Quote: "Protect AI applications against rapidly evolving threats, including prompt injections, denial of service, and data leakage."Source: [cisco.com](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html), read 2026-11-16. - **cisco-04** (vendor): Cisco says AI Defense enforces at the network level, with no agents or libraries needed on the application side.Quote: "Cisco enforces AI security at the network-level without the need for agents or libraries."Source: [cisco.com](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html), read 2026-11-16. - **cisco-05** (vendor): The AI Defense Intercept plugin for Kong Gateway inspects AI/LLM and MCP requests and responses and blocks or forwards traffic on the AI Defense verdict.Quote: "It blocks or forwards traffic based on the AI Defense verdict."Source: [github.com](https://github.com/cisco-ai-defense/ai-defense-kong-plugin), read 2026-11-16. - **cisco-06** (vendor): The Kong plugin documents a callout timeout setting (default 10 seconds, range 100 to 60000 ms) for each synchronous inspection.Quote: "Inspection callouts are synchronous, bounded by `timeout` (default 10s)."Source: [github.com](https://github.com/cisco-ai-defense/ai-defense-kong-plugin), read 2026-11-16. - **cisco-07** (vendor): The Kong plugin fails closed by default: traffic is blocked when inspection is unavailable, and an operator can set it to fail open.Quote: "Block traffic when inspection is unavailable."Source: [github.com](https://github.com/cisco-ai-defense/ai-defense-kong-plugin), read 2026-11-16. - **cisco-08** (none-found): None found on 2026-11-16: a published latency figure for an inline AI Defense decision on the AI Defense product page or the Kong plugin and Python SDK READMEs.Source: [cisco.com](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html), read 2026-11-16. - **cisco-09** (vendor): Cisco says DefenseClaw endpoint hooks fail closed by default, with Cisco Secure Client deployments defaulting to fail open when no policy is specified.Quote: "Only Cisco Secure Client deployments default hook delivery to fail-open when no policy is specified"Source: [github.com](https://github.com/cisco-ai-defense/defenseclaw), read 2026-11-16. - **cisco-10** (vendor): DefenseClaw documentation says its guardrail checks a prompt, a shell command or a tool call before the agent acts and answers allow, alert, ask or block.Quote: "The guardrail checks what your agent is about to do (a prompt, a shell command, a tool call) against rules, and answers allow, alert, ask you, or block."Source: [cisco-ai-defense.github.io](https://cisco-ai-defense.github.io/defenseclaw/docs/guardrail/), read 2026-11-16. - **cisco-11** (vendor): DefenseClaw documentation says support on Linux, macOS and Windows is marked Supported only after a live test verified hooks and a block before the tool call ran.Quote: "means a live test on that OS verified the agent's hooks and a block before the tool call ran."Source: [cisco-ai-defense.github.io](https://cisco-ai-defense.github.io/defenseclaw/docs/support-matrix/), read 2026-11-16. - **cisco-12** (vendor): The DefenseClaw compatibility table lists hook contracts for Claude Code that inspect prompts, tool calls and tool results, with similar rows for Codex, Cursor, GitHub Copilot CLI and others.Quote: "Claude Code | `>=2.1.154, <2.1.219` | `claudecode-hooks-v1` / `v7` | 28 | prompt, tool_call, tool_result, event_content"Source: [cisco-ai-defense.github.io](https://cisco-ai-defense.github.io/defenseclaw/docs/connectors/compatibility/), read 2026-11-16. - **cisco-13** (vendor): On 2026-10-21, Cisco said DefenseClaw checks every prompt, tool call and tool result in an NVIDIA OpenShell sandbox using its own rules and Cisco AI Defense.Quote: "Every prompt, tool call and tool result is checked and observed by DefenseClaw’s rules and by Cisco AI Defense."Source: [blogs.cisco.com](https://blogs.cisco.com/ai/protect-enterprise-coding-agents-with-cisco-defenseclaw-and-nvidia-openshell), read 2026-11-16. - **cisco-14** (vendor): On 2026-10-21, Cisco said DefenseClaw decides each web destination for a sandboxed coding agent by host name without decrypting traffic, which covers outbound network calls.Quote: "DefenseClaw decides each web destination by host name, without decrypting the traffic."Source: [blogs.cisco.com](https://blogs.cisco.com/ai/protect-enterprise-coding-agents-with-cisco-defenseclaw-and-nvidia-openshell), read 2026-11-16. - **cisco-15** (vendor): On 2026-10-21, Cisco described the sandbox as supplying provider credentials only for authorized requests, so the agent sees placeholders instead of real credentials.Quote: "Provider credentials are supplied only for authorized requests; inside the sandbox, the agent sees placeholders."Source: [blogs.cisco.com](https://blogs.cisco.com/ai/protect-enterprise-coding-agents-with-cisco-defenseclaw-and-nvidia-openshell), read 2026-11-16. - **cisco-17** (vendor): On 2026-10-21, Cisco said every block, approval and decision is recorded with the agent, sandbox and policy behind it and can be sent to Splunk or any OpenTelemetry backend.Quote: "Every block, approval and decision is recorded along with the agent, the sandbox and the policy behind it, and can be sent to Splunk or any OpenTelemetry backend."Source: [blogs.cisco.com](https://blogs.cisco.com/ai/protect-enterprise-coding-agents-with-cisco-defenseclaw-and-nvidia-openshell), read 2026-11-16. - **cisco-18** (vendor): DefenseClaw observability documentation lists OTLP, Splunk, Galileo, file and webhook destinations for logs, traces and metrics.Quote: "goes to the destinations you add: an OpenTelemetry (OTLP) collector, Splunk, Galileo, a file, or a webhook."Source: [cisco-ai-defense.github.io](https://cisco-ai-defense.github.io/defenseclaw/docs/observability/), read 2026-11-16. - **cisco-19** (vendor): DefenseClaw documentation says human approval is off by default, with a default minimum severity of HIGH when turned on.Quote: "Human approval is off by default. Its default minimum severity is HIGH."Source: [cisco-ai-defense.github.io](https://cisco-ai-defense.github.io/defenseclaw/docs/guardrail/), read 2026-11-16. - **cisco-20** (vendor): DefenseClaw documentation states a limit: its command-text check cannot see what a child program writes after a command starts, so protected files need a sandbox or file permissions.Quote: "It cannot see what a child program writes after a command starts. For protected files, add a sandbox or file permissions."Source: [cisco-ai-defense.github.io](https://cisco-ai-defense.github.io/defenseclaw/docs/guardrail/), read 2026-11-16. - **cisco-27** (vendor): The MCP Scanner runs as a stand-alone command-line tool or a REST API server, installed with uv.Quote: "Run scanner as a stand-alone CLI tool or REST API server"Source: [github.com](https://github.com/cisco-ai-defense/mcp-scanner), read 2026-11-16. - **cisco-29** (vendor): DefenseClaw documentation says its public website is the source of truth for install, configuration and operator workflows, and the code is Apache-2.0.Quote: "is the source of truth for installation, setup, configuration, commands, and operator workflows"Source: [github.com](https://github.com/cisco-ai-defense/defenseclaw), read 2026-11-16. - **cisco-33** (third-party): SiliconANGLE reported developers can install DefenseClaw in about five minutes and that it scans MCP tools and plug-ins an agent uses.Quote: "Cisco says developers can install DefenseClaw in about five minutes."Source: [siliconangle.com](https://siliconangle.com/2026/03/23/cisco-debuts-new-ai-agent-security-features-open-source-defenseclaw-tool/), read 2026-11-16. - **cisco-35** (vendor): Cisco says DefenseClaw streams every scan result, block decision, tool call and alert as structured events.The exact quote is on the source page.Source: [blogs.cisco.com](https://blogs.cisco.com/ai/defenseclaw-is-live), read 2026-11-16. - **cisco-37** (third-party): SiliconANGLE reported DefenseClaw sends telemetry on the risks it finds to Splunk through a prepackaged connector.Quote: "DefenseClaw sends telemetry about the cybersecurity risks that it finds to Splunk through a prepackaged connector."Source: [siliconangle.com](https://siliconangle.com/2026/03/23/cisco-debuts-new-ai-agent-security-features-open-source-defenseclaw-tool/), read 2026-11-16. - **cisco-38** (vendor): DefenseClaw documentation describes an enterprise edition that an administrator deploys with a package or an MDM and that users cannot turn off.Quote: "the managed standalone profile an administrator deploys with a package or an MDM. Users can't turn it off."Source: [cisco-ai-defense.github.io](https://cisco-ai-defense.github.io/defenseclaw/docs/support-matrix/), read 2026-11-16. - **cisco-41** (vendor): Cisco says its planned MCP gateway applies authorization policies across tools and blocks out-of-bounds agent actions, but also says the features are in development.Quote: "the MCP gateway applies authorization policies consistently across tools and services."Source: [blogs.cisco.com](https://blogs.cisco.com/security/security-agentic-ai-how-cisco-brings-zero-trust-to-your-new-digital-workforce), read 2026-11-16. - **cisco-44** (vendor): The AI Defense Python SDK README says it can inspect AWS AgentCore and Strands event streams, releasing events only after server acknowledgement.Quote: "Inspect AgentCore, Strands, and adapter-based provider streams, releasing events only after server acknowledgement."Source: [github.com](https://github.com/cisco-ai-defense/ai-defense-python-sdk), read 2026-11-16. - **cisco-45** (observed): The GitHub org lists a Cisco AI Defense plugin for Google ADK that does callback-based LLM and MCP inspection, plus LangChain and LiteLLM integrations.The exact quote is on the source page.Source: [api.github.com](https://api.github.com/orgs/cisco-ai-defense/repos), read 2026-11-16. - **cisco-48** (vendor): Cisco AI Access, part of Secure Access SSE, inspects web traffic to identify shadow AI use; it is network-based rather than a browser-agent control.Quote: "AI Access inspects web traffic to identify uses of shadow AI across your organization."Source: [cisco.com](https://www.cisco.com/site/us/en/products/security/secure-access/ai-access/index.html), read 2026-11-16. - **cisco-49** (vendor): Cisco AI Access can block or redirect requests to approved AI services and enforce data-loss prevention, which covers pasting data into third-party AI tools.Quote: "blocking or redirecting requests to approved services"Source: [cisco.com](https://www.cisco.com/site/us/en/products/security/secure-access/ai-access/index.html), read 2026-11-16. - **cisco-50** (vendor): The AI Defense Python SDK has a per-call mode setting where enforce blocks, monitor logs only and off disables inspection.Quote: ""mode": "enforce", # "monitor" to log only, "enforce" to block, "off" to disable"Source: [github.com](https://github.com/cisco-ai-defense/ai-defense-python-sdk), read 2026-11-16. - **cisco-51** (vendor): The SDK README shows inspecting a tool response for data leakage such as PII, PCI and PHI.Quote: "Inspect a tool response for data leakage (PII, PCI, PHI)"Source: [github.com](https://github.com/cisco-ai-defense/ai-defense-python-sdk), read 2026-11-16. - **cisco-52** (none-found): None found on 2026-11-16: a statement that AI Defense or DefenseClaw blocks a package install; the MCP Scanner scans PyPI and npm packages on request but no inline install block is documented in pages read.Source: [github.com](https://github.com/cisco-ai-defense/mcp-scanner), read 2026-11-16. - **cisco-53** (vendor): On 2026-10-28, Cisco said its Project Quarry detection language is in production in AI Defense and maps findings to alert, block, redact, quarantine or log.Quote: "Call Quarry from the application path and map findings to alert, block, redact, quarantine, or log."Source: [blogs.cisco.com](https://blogs.cisco.com/ai/project-quarry-agent-security-detections), read 2026-11-16. - **cisco-54** (vendor): On 2026-11-09, Cisco published a blog titled 'Cisco AI Defense to Add Anthropic's Claude Mythos for AI Red Teaming'.Quote: "Cisco AI Defense to Add Anthropic’s Claude Mythos for AI Red Teaming"Source: [blogs.cisco.com](https://blogs.cisco.com/ai/project-quarry-agent-security-detections), read 2026-11-16. - **cisco-56** (none-found): None found on 2026-11-16: a published price for Cisco AI Defense on its product page; the free Explorer Edition is announced but no paid list price was found.Source: [cisco.com](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html), read 2026-11-16. - **cisco-57** (third-party): CRN reported that at RSAC 2026 Cisco introduced a free Explorer Edition of AI Defense for red teaming agents.Quote: "Cisco also debuted updates to its AI Defense offering with the introduction of a free “Explorer Edition.”"Source: [crn.com](https://www.crn.com/news/security/2026/cisco-unveils-zero-trust-for-ai-agents-5-things-to-know), read 2026-11-16. - **cisco-58** (vendor): The AI Defense product page quotes Kent Noyes of World Wide Technology, a named customer voice.Quote: "The adoption of AI exposes companies to new risks that traditional cybersecurity solutions don't address."Source: [cisco.com](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html), read 2026-11-16. - **cisco-59** (none-found): None found on 2026-11-16: independent analyst coverage (Gartner, Forrester, IDC) or a funding round for AI Defense on the pages read; Cisco is a public company.Source: [cisco.com](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html), read 2026-11-16. ## News mentioning Cisco - [Cisco open sources Project Quarry agent detections](/news/cisco-project-quarry-agent-detections/) - [Cisco DefenseClaw pairs with NVIDIA OpenShell sandbox](/news/cisco-defenseclaw-openshell-coding-agents/) ## Related pages - [All vendor profiles](/vendors/) - [Check Point profile](/vendors/check-point/) - [Cloudflare profile](/vendors/cloudflare/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # Cloudflare AI runtime security URL: https://airuntimesecuritytools.com/vendors/cloudflare/ Updated: 2026-11-16 Cloudflare enforces at the edge, in front of the application. AI Security for Apps scores each prompt for injection so customers can block it with WAF rules, AI Gateway Guardrails can flag or block prompts before the model, and AI Gateway DLP discards the provider response and returns a status 400 error on a block. Pricing terms are public: core AI Gateway features and DLP are free, and guardrails are billed as Workers AI tokens. Gaps from 16 Nov 2026: blocking of shell, file, package or credential actions on endpoints, inline latency and fail-open or fail-closed behavior are none found, and the detection fields are an unpriced Enterprise add-on. ## Strength and weakness for AI Runtime Security Tools **Strength:** Prompt injection scoring with block rules at the edge plus Guardrails and DLP that can block prompts and responses across model providers (cloudflare-01, cloudflare-05, cloudflare-08). **Weakness:** None found on 2026-11-16 for endpoint actions, inline latency or fail-open/closed behavior, and detection fields are an Enterprise paid add-on with no public price (cloudflare-11, cloudflare-12, cloudflare-15, cloudflare-24). ## Scores and evidence Cloudflare is 13th in the overall ranking at 46.0 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Cloudflare: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Action types stopped before they complete | 5 of 10 | 30 | Prompt injection blocking by WAF rule, hazardous output blocking and DLP blocking of prompts, responses and tool call text are documented. On-machine action blocking: none found. | [cloudflare-02](#fact-cloudflare-02), [cloudflare-05](#fact-cloudflare-05), [cloudflare-06](#fact-cloudflare-06), [cloudflare-08](#fact-cloudflare-08), [cloudflare-09](#fact-cloudflare-09), [cloudflare-17](#fact-cloudflare-17), [cloudflare-24](#fact-cloudflare-24) | | Layers covered | 5 of 10 | 20 | The application and model layers are covered by an edge WAF and a proxy, with partial agent coverage through MCP portals. Endpoint: none found. | [cloudflare-03](#fact-cloudflare-03), [cloudflare-07](#fact-cloudflare-07), [cloudflare-17](#fact-cloudflare-17), [cloudflare-18](#fact-cloudflare-18) | | Latency and failure mode | 0 of 10 | 15 | Only buffering behavior for streamed responses is documented. A latency figure and fail-open or fail-closed behavior: none found. | [cloudflare-11](#fact-cloudflare-11), [cloudflare-12](#fact-cloudflare-12), [cloudflare-10](#fact-cloudflare-10) | | Deployment | 6 of 10 | 15 | Deployment needs the WAF on the zone and labeled endpoints, or a proxy with one line of code. A time to deploy in minutes or days: none found. | [cloudflare-03](#fact-cloudflare-03), [cloudflare-07](#fact-cloudflare-07), [cloudflare-13](#fact-cloudflare-13), [cloudflare-15](#fact-cloudflare-15) | | Detection and response | 6 of 10 | 5 | Guardrail logs, Logpush export and MCP portal request logs are documented. Production mode logs only metadata. | [cloudflare-06](#fact-cloudflare-06), [cloudflare-16](#fact-cloudflare-16), [cloudflare-18](#fact-cloudflare-18), [cloudflare-04](#fact-cloudflare-04) | | Transparency | 6 of 10 | 10 | Docs are cited openly and pricing terms are published (free core, token-billed guardrails), but the enterprise add-on is unpriced. Published research: none found. | [cloudflare-13](#fact-cloudflare-13), [cloudflare-14](#fact-cloudflare-14), [cloudflare-15](#fact-cloudflare-15), [cloudflare-03](#fact-cloudflare-03) | | Proof | 6 of 10 | 5 | A public company with third-party financial coverage and an observed open-source repo. Named customers or analyst coverage for these AI products: none found. | [cloudflare-19](#fact-cloudflare-19), [cloudflare-21](#fact-cloudflare-21), [cloudflare-26](#fact-cloudflare-26), [cloudflare-22](#fact-cloudflare-22), [cloudflare-23](#fact-cloudflare-23) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 13th | 46.0 | | [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) | 13th | 48.5 | | [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) | 14th | 40.5 | | [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) | joint 14th | 51.0 | | [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) | 13th | 50.0 | | [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) | joint 12th | 50.0 | ## Product, deployment, price and owner - Product assessed: AI Security for Apps (formerly Firewall for AI), AI Gateway Guardrails and DLP, and Cloudflare One MCP server portals - Deployment: Edge WAF in front of the application (AI Security for Apps) and API proxy (AI Gateway); no endpoint agent for these products found - Time to deploy: AI Gateway needs one line of code per Cloudflare docs; otherwise none found on 2026-11-16 - Price: AI Gateway core features free, DLP free, Guardrails billed as Workers AI tokens; AI Security for Apps detections are a paid Enterprise add-on with no public price found on 2026-11-16 - Owner: Cloudflare, Inc. (publicly traded, NYSE: NET) - Website: cloudflare.com ## Facts and sources - **cloudflare-01** (vendor): Cloudflare says AI Security for Apps (formerly Firewall for AI) detects prompt injection, scoring each prompt so customers can block it with WAF rules.Quote: "prompts intentionally designed to subvert the intended behavior of your LLM as specified by the developer"Source: [developers.cloudflare.com](https://developers.cloudflare.com/waf/detections/ai-security-for-apps/prompt-injection/), read 2026-11-16. - **cloudflare-02** (vendor): Cloudflare documents prompt injection scores from 1 to 99 (lower is riskier) and lets customers pick a threshold with a Block action, for example blocking below 20.Quote: "Prompt injection detection uses a score-based system rather than a binary detected/not-detected result."Source: [developers.cloudflare.com](https://developers.cloudflare.com/waf/detections/ai-security-for-apps/prompt-injection/), read 2026-11-16. - **cloudflare-03** (vendor): Cloudflare says AI Security for Apps scans requests to endpoints labeled cf-llm, handles only JSON request bodies, and needs the WAF enabled on the zone.Quote: "Currently, the detection only handles requests with a JSON content type"Source: [developers.cloudflare.com](https://developers.cloudflare.com/waf/detections/ai-security-for-apps/), read 2026-11-16. - **cloudflare-04** (vendor): Cloudflare says production mode logs only request metadata and offers no response logging, while log mode logs full request bodies using three fixed rules.The exact quote is on the source page.Source: [developers.cloudflare.com](https://developers.cloudflare.com/waf/detections/ai-security-for-apps/log-mode-vs-production-mode/), read 2026-11-16. - **cloudflare-05** (vendor): Cloudflare says AI Gateway Guardrails evaluate user prompts and can flag or block them before they reach the model.Quote: "Based on your settings, prompts can be flagged or blocked before reaching the model."Source: [developers.cloudflare.com](https://developers.cloudflare.com/ai-gateway/features/guardrails/), read 2026-11-16. - **cloudflare-06** (vendor): Cloudflare says Guardrails also inspect model responses, flag or block hazardous content, and keep logs of prompts, responses and enforced guardrails.Quote: "If hazardous content is detected, it can be flagged or blocked before being delivered to the user."Source: [developers.cloudflare.com](https://developers.cloudflare.com/ai-gateway/features/guardrails/), read 2026-11-16. - **cloudflare-07** (vendor): Cloudflare says AI Gateway Guardrails work as a proxy between the application and model providers such as OpenAI and Anthropic.Quote: "Acting as a proxy between your application and"Source: [developers.cloudflare.com](https://developers.cloudflare.com/ai-gateway/features/guardrails/), read 2026-11-16. - **cloudflare-08** (vendor): Cloudflare says AI Gateway DLP scans prompts and responses and, on a Block outcome, discards the provider response and returns a status 400 DLP error.Quote: "The provider response is discarded and replaced with a DLP error response (status 400 )."Source: [developers.cloudflare.com](https://developers.cloudflare.com/ai-gateway/features/dlp/), read 2026-11-16. - **cloudflare-09** (vendor): Cloudflare says AI Gateway DLP scans the text of tool call arguments and results when they appear in the JSON payload, but does not decode base64 content.Quote: "which includes tool call arguments and results if they appear in the JSON request or response payload"Source: [developers.cloudflare.com](https://developers.cloudflare.com/ai-gateway/features/dlp/), read 2026-11-16. - **cloudflare-10** (vendor): Cloudflare says that with DLP response scanning, streamed responses are fully buffered first, so time-to-first-token grows with full response generation time.Quote: "Time-to-first-token latency increases proportionally to the full response generation time."Source: [developers.cloudflare.com](https://developers.cloudflare.com/ai-gateway/features/dlp/), read 2026-11-16. - **cloudflare-11** (none-found): None found on 2026-11-16: published millisecond latency for an inline AI Security for Apps, Guardrails or DLP decision, looked for on the AI Security for Apps overview page.Source: [developers.cloudflare.com](https://developers.cloudflare.com/waf/detections/ai-security-for-apps/), read 2026-11-16. - **cloudflare-12** (none-found): None found on 2026-11-16: published fail-open or fail-closed behavior and timeouts when the AI Security for Apps or Guardrails check fails, looked for on the Guardrails overview page.Source: [developers.cloudflare.com](https://developers.cloudflare.com/ai-gateway/features/guardrails/), read 2026-11-16. - **cloudflare-13** (vendor): Cloudflare says AI Gateway is available on all plans and core features are free, needing a Cloudflare account and one line of code.Quote: "AI Gateway is available to use on all plans."Source: [developers.cloudflare.com](https://developers.cloudflare.com/ai-gateway/reference/pricing/), read 2026-11-16. - **cloudflare-14** (vendor): Cloudflare says DLP scanning in AI Gateway is free on all plans, and Guardrails are billed as Workers AI token-based inference.Quote: "DLP scanning in AI Gateway is free on all plans."Source: [developers.cloudflare.com](https://developers.cloudflare.com/ai-gateway/reference/pricing/), read 2026-11-16. - **cloudflare-15** (vendor): Cloudflare lists AI Security for Apps by plan: endpoint discovery on all plans, but detection fields and the log-mode ruleset only as a paid Enterprise add-on with no public price.Quote: "PII detection, prompt injection scoring, unsafe topic detection, custom topics"Source: [developers.cloudflare.com](https://developers.cloudflare.com/waf/detections/ai-security-for-apps/), read 2026-11-16. - **cloudflare-16** (vendor): Cloudflare says AI Gateway Logpush is available to all AI Gateway customers, billed by Logpush pricing on exported log volume.Quote: "is available to all AI Gateway customers"Source: [developers.cloudflare.com](https://developers.cloudflare.com/ai-gateway/reference/pricing/), read 2026-11-16. - **cloudflare-17** (vendor): Cloudflare says MCP server portals let admins curate which tools and prompt templates users reach through one Access-protected endpoint.Quote: "Admins can tailor an MCP portal to a particular use case by choosing the specific tools and prompt templates"Source: [developers.cloudflare.com](https://developers.cloudflare.com/cloudflare-one/access-controls/ai-controls/mcp-portals/), read 2026-11-16. - **cloudflare-18** (vendor): Cloudflare says Access logs individual tool requests through an MCP portal and portal traffic can optionally route through Cloudflare Gateway for HTTP logging and DLP scanning (page last updated Nov 5, 2026).Quote: "Cloudflare Access logs the individual requests made using the tools in the portal."Source: [developers.cloudflare.com](https://developers.cloudflare.com/cloudflare-one/access-controls/ai-controls/mcp-portals/), read 2026-11-16. - **cloudflare-19** (vendor): Cloudflare is a publicly traded company that debuted on the New York Stock Exchange (NYSE: NET) on 13 September 2019.Quote: "Cloudflare made its debut on the public market as it rang the bell at the New York Stock Exchange"Source: [cloudflare.com](https://www.cloudflare.com/about/), read 2026-11-16. - **cloudflare-20** (vendor): On 2026-11-05, Cloudflare announced its third quarter 2026 financial results on its press page.Quote: "Cloudflare Announces Third Quarter 2026 Financial Results"Source: [cloudflare.com](https://www.cloudflare.com/press/), read 2026-11-16. - **cloudflare-21** (third-party): On the 2026 Q3 earnings day, SiliconANGLE reported Cloudflare shares surged as revenue growth accelerated to 39%.Quote: "Cloudflare shares surge as revenue growth accelerates to 39% and outlook rises"Source: [siliconangle.com](https://siliconangle.com/?s=cloudflare+firewall+for+ai), read 2026-11-16. - **cloudflare-22** (observed): The Cloudflare MCP servers repository on GitHub is Apache-2.0 licensed with about 4.5k stars, last pushed 2026-10-22.Quote: "Apache-2.0 license"Source: [github.com](https://github.com/cloudflare/mcp-server-cloudflare), read 2026-11-16. - **cloudflare-23** (none-found): None found on 2026-11-16: named customers or analyst coverage specific to AI Security for Apps, AI Gateway Guardrails or MCP portals, looked for on the press page.Source: [cloudflare.com](https://www.cloudflare.com/press/), read 2026-11-16. - **cloudflare-24** (none-found): None found on 2026-11-16: documented blocking of shell commands, file reads or writes, package installs or credential access on endpoints by these AI products, looked for on the AI Security for Apps and AI Gateway docs.Source: [developers.cloudflare.com](https://developers.cloudflare.com/ai-gateway/), read 2026-11-16. - **cloudflare-25** (vendor): Cloudflare says it serves about 102 million HTTP requests per second and runs sixty-plus services across SASE, application security and full-stack development.Quote: "We serve an average of 102 million HTTP requests per second"Source: [cloudflare.com](https://www.cloudflare.com/about/), read 2026-11-16. - **cloudflare-26** (third-party): Wikipedia says Cloudflare is headquartered in San Francisco and was founded on July 26, 2009 by Matthew Prince, Lee Holloway and Michelle Zatlyn.Quote: "founded on July 26, 2009, by Matthew Prince"Source: [en.wikipedia.org](https://en.wikipedia.org/wiki/Cloudflare), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [Cisco profile](/vendors/cisco/) - [CrowdStrike profile](/vendors/crowdstrike/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # CrowdStrike AI runtime security URL: https://airuntimesecuritytools.com/vendors/crowdstrike/ Updated: 2026-11-16 CrowdStrike enforces at four layers through products it already sells. Falcon Guardian, the successor to Falcon AI Detection and Response, uses agent-level hooks for Claude Code, Codex, GitHub Copilot, OpenCode and Cursor to block malicious activity before it runs. The Falcon sensor blocks malicious open-source packages as they reach the endpoint, and Falcon AI Gateway inspects prompts and responses on routed traffic. Falcon Complete for Guardian adds 24/7 managed response. Gaps from 16 Nov 2026: CrowdStrike's own latency figure, failure mode and credential access blocking as a named action are none found, the gateway acts only on routed traffic, and cloud workload coverage is pre-beta. ## Strength and weakness for AI Runtime Security Tools **Strength:** Enforces at four layers: model and application via the gateway, agent via hooks, endpoint via the sensor, and blocks malicious packages (facts 10, 16, 18). **Weakness:** None found on 2026-11-16 for published latency, failure mode, or credential access blocking (facts 57, 62). ## Scores and evidence CrowdStrike is 10th in the overall ranking at 52.0 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **CrowdStrike: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Action types stopped before they complete | 6 of 10 | 30 | The vendor claims blocking of malicious packages, prompt injection, sensitive data exposure and unnamed harmful activity through agent hooks; paste and upload control is in a separate product. Credential access as a named action: none found. | [crowdstrike-16](#fact-crowdstrike-16), [crowdstrike-10](#fact-crowdstrike-10), [crowdstrike-13](#fact-crowdstrike-13), [crowdstrike-18](#fact-crowdstrike-18), [crowdstrike-52](#fact-crowdstrike-52), [crowdstrike-36](#fact-crowdstrike-36), [crowdstrike-62](#fact-crowdstrike-62) | | Layers covered | 7 of 10 | 20 | The vendor claims endpoint sensor, agent hooks and a gateway for prompts and responses. The gateway acts only on routed traffic, and cloud workload coverage is pre-beta. | [crowdstrike-09](#fact-crowdstrike-09), [crowdstrike-10](#fact-crowdstrike-10), [crowdstrike-18](#fact-crowdstrike-18), [crowdstrike-19](#fact-crowdstrike-19), [crowdstrike-27](#fact-crowdstrike-27) | | Latency and failure mode | 0 of 10 | 15 | The only latency figure belongs to a company the vendor is acquiring. The vendor's own figure and failure mode: none found. | [crowdstrike-57](#fact-crowdstrike-57), [crowdstrike-22](#fact-crowdstrike-22), [crowdstrike-08](#fact-crowdstrike-08) | | Deployment | 6 of 10 | 15 | Deployment uses the existing sensor, MDM and a base-URL switch on Windows, macOS and Linux, and an API integration is observed on GitHub. Time to deploy and public docs: none found. | [crowdstrike-02](#fact-crowdstrike-02), [crowdstrike-17](#fact-crowdstrike-17), [crowdstrike-21](#fact-crowdstrike-21), [crowdstrike-30](#fact-crowdstrike-30), [crowdstrike-51](#fact-crowdstrike-51), [crowdstrike-60](#fact-crowdstrike-60) | | Detection and response | 7 of 10 | 5 | The vendor claims SIEM export, causal investigation, managed detection and response, and automatic containment. All of it is vendor-only. | [crowdstrike-38](#fact-crowdstrike-38), [crowdstrike-40](#fact-crowdstrike-40), [crowdstrike-42](#fact-crowdstrike-42), [crowdstrike-39](#fact-crowdstrike-39), [crowdstrike-11](#fact-crowdstrike-11) | | Transparency | 4 of 10 | 10 | Research is published, and endpoint bundle prices are public, but a price for these AI products and public product docs: none found. | [crowdstrike-45](#fact-crowdstrike-45), [crowdstrike-23](#fact-crowdstrike-23), [crowdstrike-48](#fact-crowdstrike-48), [crowdstrike-49](#fact-crowdstrike-49), [crowdstrike-46](#fact-crowdstrike-46), [crowdstrike-47](#fact-crowdstrike-47) | | Proof | 7 of 10 | 5 | Third-party coverage of acquisitions and a vendor customer story. Analyst coverage specific to AI security: none found. | [crowdstrike-05](#fact-crowdstrike-05), [crowdstrike-43](#fact-crowdstrike-43), [crowdstrike-44](#fact-crowdstrike-44) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 10th | 52.0 | | [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) | 9th | 56.5 | | [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) | 10th | 47.0 | | [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) | joint 8th | 60.5 | | [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) | 8th | 58.0 | | [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) | 10th | 55.5 | ## Product, deployment, price and owner - Product assessed: Falcon Guardian (sensor, agent hooks, SafeMind detection), Falcon AI Gateway, Kill Switch, and Falcon sensor package protection - Deployment: Falcon sensor on endpoints with agent-level hooks, a routed Falcon AI Gateway for apps and workloads beyond the sensor, and device-management rollout for the Cowork integration - Time to deploy: none found on 2026-11-16 - Price: none found on 2026-11-16 for Falcon Guardian; endpoint bundles are priced (Falcon Go $7.99 per device monthly) and Falcon AI Gateway is included with Guardian at no additional cost - Owner: CrowdStrike Holdings, Inc. (public company; CrowdStrike sells the products itself) - Website: crowdstrike.com ## Facts and sources - **crowdstrike-01** (vendor): CrowdStrike says Falcon Guardian discovers shadow AI, agents, applications, models, MCP servers, skills and tools on endpoints.Quote: "Discover shadow AI, agents, applications, models, MCP servers, skills, and tools."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/platform/falcon-guardian-aidr/), read 2026-11-16. - **crowdstrike-02** (vendor): On 2026-09-01 CrowdStrike introduced Falcon Guardian, the successor to Falcon AI Detection and Response (AIDR), and says it discovers AI agents on supported Windows, macOS and Linux endpoints.Quote: "Falcon Guardian continuously discovers known and previously unknown AI agents across supported Windows, macOS, and Linux endpoints."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/), read 2026-11-16. - **crowdstrike-03** (vendor): CrowdStrike names Falcon Guardian as the evolution of Falcon AI Detection and Response (AIDR) and its flagship AI security product.Quote: "the evolution of Falcon AI Detection and Response (AIDR) and CrowdStrike’s flagship solution for the AIDR category"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/), read 2026-11-16. - **crowdstrike-04** (vendor): CrowdStrike lists George Kurtz as CEO and Founder and Michael Sentonas as President on its executive team page.Quote: "George Kurtz CEO & Founder Michael Sentonas President"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/about-us/executive-team/), read 2026-11-16. - **crowdstrike-05** (third-party): SiliconANGLE reported on 2026-11-03 that CrowdStrike Holdings plans to buy Fiddler AI and Dreadnode, two AI agent security startups, announced with a Falcon Guardian update.Quote: "CrowdStrike Holdings Inc. today disclosed plans to buy Fiddler AI Inc. and Dreadnode Inc., two startups that specialize in securing artificial intelligence agents."Source: [siliconangle.com](https://siliconangle.com/2026/11/03/crowdstrike-to-acquire-cybersecurity-startups-fiddler-ai-dreadnode-for-reported-495m/), read 2026-11-16. - **crowdstrike-06** (vendor): On 2026-11-03 CrowdStrike announced its intent to acquire Fiddler, to become Falcon Observe for AI agent observability, generally available in Q4 FY27 (a future product, not scored as available).Quote: "Fiddler will become part of the Falcon platform as Falcon Observe, generally available in Q4 FY27."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-to-acquire-fiddler/), read 2026-11-16. - **crowdstrike-08** (third-party): SiliconANGLE reports that Fiddler, which CrowdStrike plans to acquire, says its Centor models add under 80 milliseconds to reviewed requests; this is Fiddler's figure and not a published Falcon Guardian latency.Quote: "According to the company, Centor adds less than 80 milliseconds of latency to the requests it reviews."Source: [siliconangle.com](https://siliconangle.com/2026/11/03/crowdstrike-to-acquire-cybersecurity-startups-fiddler-ai-dreadnode-for-reported-495m/), read 2026-11-16. - **crowdstrike-09** (vendor): On 2026-11-03 CrowdStrike said the Falcon sensor gives OS-level visibility into agent processes, commands and resource access, available now.Quote: "The Falcon sensor provides deep visibility into agents’ activity, including the processes they launch, commands they run, and resources they access."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-advances-ai-security-and-safety-falcon-guardian/), read 2026-11-16. - **crowdstrike-10** (vendor): On 2026-11-03 CrowdStrike said agent-level hooks for Claude Code, OpenAI Codex, GitHub Copilot, OpenCode and Cursor inspect prompts and block malicious activity before harmful actions run (hooks act on the agent, not specifically on MCP tool calls).Quote: "Hooks for Claude Code, OpenAI Codex, GitHub Copilot, OpenCode, and Cursor enable organizations to inspect prompts and block malicious activity at the source, before harmful actions are executed."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-advances-ai-security-and-safety-falcon-guardian/), read 2026-11-16. - **crowdstrike-11** (vendor): On 2026-09-01 CrowdStrike said Falcon Guardian can drive automatic containment by blocking malicious agent behaviors and compromised assets at runtime.Quote: "drive automatic containment by blocking malicious agent behaviors and compromised assets at runtime across agents"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/), read 2026-11-16. - **crowdstrike-12** (vendor): On 2026-09-01 CrowdStrike said Falcon Guardian lets organizations define which supported agent types may run on managed endpoints and blocks unauthorized types.Quote: "helping security teams sanction approved agents and prevent unauthorized agent types from running."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/), read 2026-11-16. - **crowdstrike-13** (vendor): CrowdStrike says Falcon Guardian protects supported AI interactions against prompt injection and sensitive data exposure.Quote: "Existing Falcon Guardian controls continue to protect supported AI interactions against threats such as prompt injection and sensitive data exposure."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/), read 2026-11-16. - **crowdstrike-14** (vendor): On 2026-11-03 CrowdStrike said Falcon Guardian threat detection is powered by SafeMind models that identify malicious intent such as prompt injection.Quote: "These models help identify malicious intent in AI interactions, including prompt injection and other attacks targeting AI systems."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-advances-ai-security-and-safety-falcon-guardian/), read 2026-11-16. - **crowdstrike-16** (vendor): On 2026-09-02 CrowdStrike said the Falcon sensor blocks malicious open-source packages as they reach the endpoint, before embedded code executes (a sensor capability, separate from Guardian).Quote: "It detects and blocks malicious open-source packages as they reach the endpoint, before any embedded code can execute"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-extends-endpoint-security-to-stop-supply-chain-attacks/), read 2026-11-16. - **crowdstrike-17** (vendor): CrowdStrike says package monitoring covers npm and PyPI downloads on Windows, macOS and Linux through the existing Falcon sensor.Quote: "constant watch over package downloads across npm and PyPI on Windows, macOS, and Linux"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-extends-endpoint-security-to-stop-supply-chain-attacks/), read 2026-11-16. - **crowdstrike-18** (vendor): On 2026-11-13 CrowdStrike said Falcon AI Gateway routes AI traffic through a central gateway that inspects prompts and responses and enforces policy on managed and unmanaged devices.Quote: "It routes AI traffic through a centralized gateway to inspect prompts and responses, monitor usage, and enforce policy across managed and unmanaged devices."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-ai-gateway-one-control-point-for-ai-traffic/), read 2026-11-16. - **crowdstrike-19** (vendor): CrowdStrike states the gateway acts only on traffic routed through it and cannot stop local processes.Quote: "Falcon AI Gateway acts on traffic routed through it and can't stop local processes."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-ai-gateway-one-control-point-for-ai-traffic/), read 2026-11-16. - **crowdstrike-20** (vendor): CrowdStrike states the gateway does not discover or redirect AI traffic automatically; teams choose which workloads to route through it.Quote: "This doesn't discover or redirect AI traffic automatically; it gives teams a deliberate control point for the workloads they choose to route through it."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-ai-gateway-one-control-point-for-ai-traffic/), read 2026-11-16. - **crowdstrike-21** (vendor): CrowdStrike says developers connect an existing SDK or coding tool to the gateway by changing the base URL and using a gateway key.Quote: "Developers connect an existing SDK or coding tool by changing the base URL to point at the gateway and using a gateway key."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-ai-gateway-one-control-point-for-ai-traffic/), read 2026-11-16. - **crowdstrike-22** (vendor): On 2026-11-13 CrowdStrike discussed keeping the gateway request path short but gave no latency number or fail-open or fail-closed behavior in that post.Quote: "we learned to keep the request path short, since every extra hop adds another failure point and more time."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-ai-gateway-one-control-point-for-ai-traffic/), read 2026-11-16. - **crowdstrike-23** (vendor): On 2026-11-03 CrowdStrike said Falcon AI Gateway is included with Falcon Guardian at no additional cost.Quote: "Included with Falcon Guardian at no additional cost, it extends AI security to applications and workloads beyond direct endpoint sensor coverage, including unmanaged environments."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-introduces-ai-kill-switch/), read 2026-11-16. - **crowdstrike-24** (vendor): On 2026-11-03 CrowdStrike said AI Kill Switch in Falcon Guardian directs connected AI gateways (Falcon or third-party) to block access to the models that power agents.Quote: "Security teams activate AI Kill Switch in Falcon Guardian, which directs connected AI gateways to block access to the AI models that power agents' decisions and actions."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-introduces-ai-kill-switch/), read 2026-11-16. - **crowdstrike-27** (vendor): On 2026-11-03 CrowdStrike described pre-beta coverage of cloud workload and Kubernetes agents, including LangChain and LangGraph apps.Quote: "In cloud workloads and Kubernetes, Falcon Guardian will extend these capabilities to enterprise-developed AI agents and applications, including those built with LangChain and LangGraph."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-advances-ai-security-and-safety-falcon-guardian/), read 2026-11-16. - **crowdstrike-30** (vendor): On 2026-11-03 CrowdStrike said admins turn on the Cowork integration through existing device management, as they deploy the Falcon sensor.Quote: "Admins will turn the integration on through their existing device management, the same way they deploy the Falcon sensor today"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-secures-ai-agents-inside-claude-cowork/), read 2026-11-16. - **crowdstrike-36** (vendor): On 2026-11-04 CrowdStrike said Seraphic can enforce copy, paste, upload and download policies to keep sensitive data from unsanctioned web apps and generative AI tools (a separate product from Guardian).Quote: "Teams can enforce policies on browser actions such as copy, paste, upload, and download; this helps prevent sensitive data from reaching unsanctioned web applications and generative AI tools."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-brings-enterprise-browser-security-to-falcon-platform/), read 2026-11-16. - **crowdstrike-38** (vendor): On 2026-09-01 CrowdStrike said Guardian exports agent telemetry into Falcon Next-Gen SIEM as first-party data.Quote: "Falcon Guardian natively exports agent telemetry into CrowdStrike Falcon® Next-Gen SIEM as first-party data, pre-mapped to its schema for immediate correlation, detection, and automation"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/), read 2026-11-16. - **crowdstrike-39** (vendor): On 2026-11-03 CrowdStrike listed Guardian additions: Falcon Data Replicator export of prompt activity, Flight Control support, and token consumption dashboards.Quote: "Falcon Data Replicator integration for prompt activity, Falcon Flight Control support, new executive summary and token consumption dashboards, and expanded Windows AI application coverage."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-advances-ai-security-and-safety-falcon-guardian/), read 2026-11-16. - **crowdstrike-40** (vendor): On 2026-09-01 CrowdStrike said Guardian reconstructs agent sessions and downstream execution into a causal investigation.Quote: "Falcon Guardian reconstructs agent sessions and downstream execution into a unified causal investigation, allowing analysts to trace suspicious activity across affected agents and systems."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/), read 2026-11-16. - **crowdstrike-42** (vendor): On 2026-11-03 CrowdStrike announced Falcon Complete for Guardian is generally available, adding 24/7 managed detection and response for AI security.Quote: "the general availability of Falcon Complete for Guardian, extending managed detection and response (MDR) services to AI security."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-advances-ai-security-and-safety-falcon-guardian/), read 2026-11-16. - **crowdstrike-43** (vendor): CrowdStrike publishes a customer story saying Cogentrix gained day-one AI visibility with Falcon Guardian (customer claim, no time figure).Quote: "See how Cogentrix gains day-one AI visibility and governs employee and agentic AI with Falcon Guardian."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/platform/falcon-guardian-aidr/), read 2026-11-16. - **crowdstrike-44** (vendor): CrowdStrike lists on its blog that it was named a Leader in the 2026 IDC MarketScape for Worldwide Augmented Identity Security Solutions (dated 2026-11-12; identity security, not AI security specifically).Quote: "CrowdStrike Named a Leader in the 2026 IDC MarketScape for Worldwide Augmented Identity Security Solutions"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-introduces-ai-kill-switch/), read 2026-11-16. - **crowdstrike-45** (vendor): CrowdStrike publishes list prices for endpoint bundles, for example Falcon Go at $7.99 per device billed monthly; Guardian is not on that page.Quote: "Falcon Go $7.99 per device, billed monthly"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/pricing/), read 2026-11-16. - **crowdstrike-46** (none-found): None found on 2026-11-16: a published price for Falcon Guardian or Falcon AIDR on the pricing page (only endpoint bundles are priced; other products say contact us).Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/pricing/), read 2026-11-16. - **crowdstrike-47** (none-found): None found on 2026-11-16: public product documentation for Falcon Guardian; the documentation address returned a Falcon login page.Source: [falcon.crowdstrike.com](https://falcon.crowdstrike.com/documentation/), read 2026-11-16. - **crowdstrike-48** (vendor): On 2026-01-09 CrowdStrike published research explaining AI tool poisoning through MCP tool descriptions and how to defend against it.Quote: "In this blog, we explain how AI tool poisoning works, the different forms it can take, and how organizations can strengthen their defenses against this type of attack."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/ai-tool-poisoning/), read 2026-11-16. - **crowdstrike-49** (vendor): CrowdStrike links an interactive Taxonomy of Prompt Injection Methods as published research from its AI agent security page.Quote: "Taxonomy of Prompt Injection Methods"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/solutions/ai-agent-security/), read 2026-11-16. - **crowdstrike-51** (observed): CrowdStrike publishes aidr-mcp-server on GitHub, which integrates with CrowdStrike AIDR APIs using an API token.Quote: "server that provides integration with CrowdStrike AIDR APIs."Source: [github.com](https://github.com/CrowdStrike/aidr-mcp-server), read 2026-11-16. - **crowdstrike-52** (observed): The CrowdStrike aidr-mcp-server README documents a prompt_guard tool that analyzes and redacts text to avoid model manipulation and malicious content.Quote: "Analyze and redact text to avoid manipulation of the model, addition of malicious content, and other undesirable data transfers."Source: [github.com](https://github.com/CrowdStrike/aidr-mcp-server), read 2026-11-16. - **crowdstrike-54** (vendor): On 2026-11-04 CrowdStrike described a Q4 roadmap feature in Next-Gen SIEM that baselines each agent and its owner and flags deviations (not yet available).Quote: "learns the normal behavior pattern for every AI agent and the person who runs it, the files it touches, the tools it calls"Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-detects-and-stops-agent-insider-threats/), read 2026-11-16. - **crowdstrike-55** (vendor): CrowdStrike says endpoint network inspection adds traffic-level controls while the Falcon sensor extends control to runtime on the endpoint.Quote: "Endpoint Network Inspection adds traffic-level controls, while the Falcon sensor extends control to runtime on the endpoint."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/platform/falcon-guardian-aidr/), read 2026-11-16. - **crowdstrike-56** (vendor): On 2026-11-03 CrowdStrike said endpoint network inspection is pre-beta and planned to expand to macOS.Quote: "CrowdStrike announced plans to expand this capability to macOS, where it will inspect traffic entering and leaving the endpoint to help security teams identify risky interactions."Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/crowdstrike-advances-ai-security-and-safety-falcon-guardian/), read 2026-11-16. - **crowdstrike-57** (none-found): None found on 2026-11-16: a published latency figure or fail-open or fail-closed behavior for Falcon AI Gateway or Guardian hooks, in the gateway blog.Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/blog/falcon-ai-gateway-one-control-point-for-ai-traffic/), read 2026-11-16. - **crowdstrike-60** (none-found): None found on 2026-11-16: a published time to deploy for Falcon Guardian on the Guardian page.Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/platform/falcon-guardian-aidr/), read 2026-11-16. - **crowdstrike-62** (none-found): None found on 2026-11-16: documented blocking of agent credential access as a named action type, on the Falcon Guardian page.Source: [crowdstrike.com](https://www.crowdstrike.com/en-us/platform/falcon-guardian-aidr/), read 2026-11-16. - **crowdstrike-63** (third-party): TechRadar reported that CrowdStrike confirmed it would acquire identity security startup SGNL for about $740 million, to support continuous access control for human, non-human and AI identities (article undated in extracted text).Quote: "CrowdStrike has confirmed it will be acquiring identity security startup SGNL for around $740 million in a mostly-cash deal, subject to regulatory approval."Source: [techradar.com](https://www.techradar.com/pro/crowdstrike-acquires-continuous-identity-firm-sgnl-for-usd740m-in-push-for-next-gen-security), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [Cloudflare profile](/vendors/cloudflare/) - [HiddenLayer profile](/vendors/hiddenlayer/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # HiddenLayer AI runtime security URL: https://airuntimesecuritytools.com/vendors/hiddenlayer/ Updated: 2026-11-16 HiddenLayer ranks first on this site, and its public documentation carries much of the score. The docs describe Agent Harness Security gating a coding agent when a prompt is submitted, before a shell command or MCP tool runs, when tool output returns and when files are read or edited. In Cursor, shell execution and file reads can be blocked before they run. Model-layer controls cover prompt injection, sensitive data and denial of service. Policy outcomes of detect, redact or block go to a webhook, Splunk or Azure Sentinel, and sessions can be replayed. Gaps: no fixed latency figure, no platform-wide fail-open or fail-closed behavior, and no price, all none found on 16 Nov 2026. ## Strength and weakness for AI Runtime Security Tools **Strength:** Agent Harness Security documents blocking before shell commands, file reads and MCP tool runs in Cursor and tool calls in Claude Code, plus model-layer prompt injection and output controls (hiddenlayer-10, hiddenlayer-12, hiddenlayer-14). **Weakness:** No published latency figure, platform-wide fail-open or fail-closed behavior, or price: none found on 2026-11-16 (hiddenlayer-24, hiddenlayer-27, hiddenlayer-32). ## Scores and evidence HiddenLayer is 1st in the overall ranking at 71.0 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **HiddenLayer: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Action types stopped before they complete | 8 of 10 | 30 | Observed docs show blocking of shell, file reads and edits, MCP tool calls, prompt injection and data leakage before they run. Some gate points are visibility only. | [hiddenlayer-10](#fact-hiddenlayer-10), [hiddenlayer-12](#fact-hiddenlayer-12), [hiddenlayer-07](#fact-hiddenlayer-07), [hiddenlayer-09](#fact-hiddenlayer-09), [hiddenlayer-14](#fact-hiddenlayer-14), [hiddenlayer-13](#fact-hiddenlayer-13) | | Layers covered | 9 of 10 | 20 | Observed docs cover model, application, agent and endpoint coding-agent gates, and third-party press confirms the coding-agent coverage. | [hiddenlayer-10](#fact-hiddenlayer-10), [hiddenlayer-11](#fact-hiddenlayer-11), [hiddenlayer-15](#fact-hiddenlayer-15), [hiddenlayer-18](#fact-hiddenlayer-18), [hiddenlayer-05](#fact-hiddenlayer-05), [hiddenlayer-41](#fact-hiddenlayer-41) | | Latency and failure mode | 4 of 10 | 15 | Docs say latency depends on hardware, and a fail-open or fail-closed timeout exists only for the code detector. A fixed latency figure and a platform-wide failure mode: none found. | [hiddenlayer-23](#fact-hiddenlayer-23), [hiddenlayer-24](#fact-hiddenlayer-24), [hiddenlayer-25](#fact-hiddenlayer-25), [hiddenlayer-26](#fact-hiddenlayer-26) | | Deployment | 6 of 10 | 15 | Observed docs detail container images, MDM rollout, OS requirements and integrations, and the feature may need enabling by the vendor. Time to deploy: none found. | [hiddenlayer-18](#fact-hiddenlayer-18), [hiddenlayer-20](#fact-hiddenlayer-20), [hiddenlayer-21](#fact-hiddenlayer-21), [hiddenlayer-19](#fact-hiddenlayer-19), [hiddenlayer-17](#fact-hiddenlayer-17), [hiddenlayer-22](#fact-hiddenlayer-22) | | Detection and response | 9 of 10 | 5 | Observed docs show policy outcomes sent to webhook, Splunk and Sentinel, recorded interactions, and replayable sessions for investigation. | [hiddenlayer-27](#fact-hiddenlayer-27), [hiddenlayer-28](#fact-hiddenlayer-28), [hiddenlayer-16](#fact-hiddenlayer-16), [hiddenlayer-29](#fact-hiddenlayer-29) | | Transparency | 6 of 10 | 10 | Public docs, release notes, research and open-source code are observed. Published price: none found. | [hiddenlayer-30](#fact-hiddenlayer-30), [hiddenlayer-33](#fact-hiddenlayer-33), [hiddenlayer-32](#fact-hiddenlayer-32), [hiddenlayer-31](#fact-hiddenlayer-31) | | Proof | 7 of 10 | 5 | Third-party press confirms funding and government contracts, and the vendor shows named CISO endorsements. Analyst coverage: none found. | [hiddenlayer-36](#fact-hiddenlayer-36), [hiddenlayer-37](#fact-hiddenlayer-37), [hiddenlayer-38](#fact-hiddenlayer-38), [hiddenlayer-34](#fact-hiddenlayer-34), [hiddenlayer-39](#fact-hiddenlayer-39) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 1st | 71.0 | | [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) | 1st | 74.5 | | [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) | 1st | 68.0 | | [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) | 1st | 77.5 | | [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) | 1st | 77.0 | | [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) | 1st | 69.5 | ## Product, deployment, price and owner - Product assessed: AI Runtime Security (AIDR and Agentic Runtime Security) plus Agent Harness Security for coding agents - Deployment: Customer-run container images, evaluation API and SDK, framework and gateway integrations, and an on-device plug-in for coding agents deployed via MDM - Time to deploy: none found on 2026-11-16 - Price: none found on 2026-11-16 - Owner: HiddenLayer, Inc. (standalone, venture-backed; no parent company found) - Website: hiddenlayer.com ## Facts and sources - **hiddenlayer-01** (vendor): HiddenLayer's Series B announcement is datelined Austin, TX, and the company describes itself as an AI security company for agentic, generative and predictive AI.The exact quote is on the source page.Source: [hiddenlayer.com](https://www.hiddenlayer.com/news/hiddenlayer-100m-series-b-ai-security), read 2026-11-16. - **hiddenlayer-02** (vendor): HiddenLayer says it was founded by security researchers who found the first AI-specific malware.Quote: "Founded by security researchers who uncovered the first-ever AI-specific malware"Source: [hiddenlayer.com](https://www.hiddenlayer.com/about-us), read 2026-11-16. - **hiddenlayer-03** (vendor): HiddenLayer's leadership page lists Christopher Sestito as Chairman, CEO and Co-Founder, with Tanner Burns, Jim Ballard and others.Quote: "Chairman of the Board, CEO & Co-Founder"Source: [hiddenlayer.com](https://www.hiddenlayer.com/about-us), read 2026-11-16. - **hiddenlayer-04** (vendor): HiddenLayer's platform has four modules: AI Discovery, AI Supply Chain Security, AI Attack Simulation and AI Runtime Security; Runtime Security is the module scored here.Quote: "AI Runtime Security"Source: [hiddenlayer.com](https://www.hiddenlayer.com/platform/ai-runtime-security), read 2026-11-16. - **hiddenlayer-05** (vendor): HiddenLayer says runtime security protects AI applications, agents and agentic workflows with runtime visibility, threat detection and inline enforcement.Quote: "HiddenLayer protects AI applications, agents, and agentic workflows with runtime visibility, threat detection, and inline enforcement."Source: [hiddenlayer.com](https://www.hiddenlayer.com/platform/ai-runtime-security), read 2026-11-16. - **hiddenlayer-06** (vendor): HiddenLayer says its inline protection can detect, block or redact unsafe actions and sensitive information according to policy and platform capabilities.Quote: "Detect, block, or redact unsafe actions and sensitive information according to policy and platform capabilities."Source: [hiddenlayer.com](https://www.hiddenlayer.com/platform/ai-runtime-security), read 2026-11-16. - **hiddenlayer-07** (vendor): HiddenLayer's agentic security page says it blocks indirect prompt injection hidden in data, documents, MCP responses or retrieved context before it affects agent execution.Quote: "Identify and block prompt injection hidden inside data, documents, MCP responses, or retrieved context before it impacts agent execution."Source: [hiddenlayer.com](https://www.hiddenlayer.com/solutions/agentic-mcp-security), read 2026-11-16. - **hiddenlayer-08** (vendor): HiddenLayer says it monitors agent actions and enforces runtime policy across APIs, MCP tools, code execution, communication tools and filesystem operations.Quote: "Monitor agent actions and enforce runtime policy across APIs, MCP tools, code execution, communication tools, and filesystem operations."Source: [hiddenlayer.com](https://www.hiddenlayer.com/solutions/agentic-mcp-security), read 2026-11-16. - **hiddenlayer-09** (vendor): HiddenLayer's guardrails page says it enforces policies against prompt injection, data leakage and sensitive information exposure in model responses, and covers PII, PHI and proprietary data.Quote: "Detect and prevent PII, PHI, and proprietary data exposure in generated content."Source: [hiddenlayer.com](https://www.hiddenlayer.com/solutions/ai-guardrails), read 2026-11-16. - **hiddenlayer-10** (observed): HiddenLayer's docs say Agent Harness Security gates coding agents when a prompt is submitted, before a tool or shell command runs, when tool output returns, and when files are read or edited.Quote: "when a prompt is submitted, before a tool or shell command runs, when tool output comes back, and when files are read or edited"Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/agent_harness/architecture.md), read 2026-11-16. - **hiddenlayer-11** (observed): HiddenLayer's docs list supported coding agents as Claude Code, Cursor and GitHub Copilot CLI, each integrated through that agent's hooks.Quote: "Claude Code hooks"Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/agent_harness/architecture.md), read 2026-11-16. - **hiddenlayer-12** (observed): HiddenLayer's docs say Cursor shell execution and file reads can be blocked before they run, and MCP execution can be blocked before the MCP tool runs.Quote: "Before a shell command runs."Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/agent_harness/architecture.md), read 2026-11-16. - **hiddenlayer-13** (observed): HiddenLayer's docs say some gate points are visibility only because they fire after the action already happened, and redaction is escalated to block where a gate cannot rewrite content.Quote: "Visibility only."Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/agent_harness/architecture.md), read 2026-11-16. - **hiddenlayer-14** (observed): HiddenLayer's docs list the findings Agentic Runtime Security surfaces: prompt injection, sensitive data, code, URLs, denial of service, off-policy language and model refusals.Quote: "Prompt injection**: attempts to override instructions or hijack the model."Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/agentic/overview.md), read 2026-11-16. - **hiddenlayer-15** (observed): HiddenLayer's docs say Agentic Runtime Security is called through evaluation endpoints, directly or with the HiddenLayer SDK, and returns findings and an action to take.Quote: "you call the evaluation endpoints (directly, with the HiddenLayer SDK) to check a model request, response, or a whole interaction"Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/agentic/overview.md), read 2026-11-16. - **hiddenlayer-16** (observed): HiddenLayer's docs say Agentic Runtime Security reconstructs the related calls of one agent run into a single replayable session for investigation.Quote: "the related calls of one agent run, reconstructed into a single, replayable conversation you can investigate and govern as a whole"Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/agentic/overview.md), read 2026-11-16. - **hiddenlayer-17** (observed): HiddenLayer's docs say Agentic Runtime Security is rolling out and may need to be enabled for a tenant by contacting a HiddenLayer representative.Quote: "Agentic Runtime Security is being rolled out and may not be enabled for your tenant yet."Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/agentic/overview.md), read 2026-11-16. - **hiddenlayer-18** (observed): HiddenLayer's docs say the runtime product is delivered as container images the customer deploys and operates in its own infrastructure.The exact quote is on the source page.Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/overview.md), read 2026-11-16. - **hiddenlayer-19** (observed): HiddenLayer lists pre-built integrations for AWS Strands, OpenAI Agents SDK, LangChain and LangGraph, Microsoft Agent Framework, Azure API Management and the TrueFoundry AI Gateway.Quote: "Azure API Management (APIM) Guardrails"Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/agentic/integrations.md), read 2026-11-16. - **hiddenlayer-20** (observed): HiddenLayer's docs say the Agent Harness plug-in is deployed to endpoints through an MDM such as Jamf or Microsoft Intune, or by copying the file, and needs a client credential and project ID.Quote: "Deploy using an MDM application, like Jamf or Microsoft Intune."Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/agent_harness/deployment.md), read 2026-11-16. - **hiddenlayer-21** (observed): HiddenLayer's docs state the Windows Intune package requires Windows 10 1809 or later or Windows 11, x64 only, and says ARM64 is not supported.Quote: "Windows 10 1809 or later, or Windows 11 (recommended), x64 only."Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/agent_harness/deployment.md), read 2026-11-16. - **hiddenlayer-22** (none-found): None found on 2026-11-16: a published time to deploy for AI Runtime Security or Agent Harness Security on the product page, the docs overview or the Agent Harness deployment page.Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/agent_harness/deployment.md), read 2026-11-16. - **hiddenlayer-23** (observed): HiddenLayer's docs say latency and throughput per Runtime Security instance depend on hardware, network and contention, and give no fixed figure.Quote: "The latency and throughput for each instance depends on many factors in the deployed environment, including underlying hardware, network conditions, and resource contention."Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/resource_requirements.md), read 2026-11-16. - **hiddenlayer-24** (observed): HiddenLayer release notes say the Interactions API now returns detection latency, so customers can measure how long each detection takes.Quote: "The Interactions API now returns detection latency, giving you visibility into how long each detection takes to process."Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/releasenotes.md), read 2026-11-16. - **hiddenlayer-25** (observed): HiddenLayer release notes say users can configure a timeout for the code detector and choose whether it fails open or closed on timeout; no platform-wide failure mode was found.Quote: "as well as the ability to configure whether the detector should fail open or closed if a timeout is encountered."Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/releasenotes.md), read 2026-11-16. - **hiddenlayer-26** (none-found): None found on 2026-11-16: a published millisecond latency figure for an inline decision, or a published fail-open or fail-closed behavior for the whole control, in the docs pages read.Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/resource_requirements.md), read 2026-11-16. - **hiddenlayer-27** (observed): HiddenLayer's docs say Runtime Policy Outcomes (detect, redact, block) can be delivered to a webhook, Splunk or Azure Sentinel destination.Quote: "A webhook or Splunk destination can subscribe to one or more data sources."Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/agentic/policy_outcomes.md), read 2026-11-16. - **hiddenlayer-28** (observed): HiddenLayer's docs say every Agent Harness interaction is recorded and visible in the console, including observe-only gate points.Quote: "Every interaction is recorded and visible in the console, including observe-only gate points"Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/agent_harness/architecture.md), read 2026-11-16. - **hiddenlayer-29** (vendor): HiddenLayer says Agent Harness Security traces agent activity from the initial prompt through execution and gives an auditable record of agent behavior.Quote: "Trace agent activity from the initial prompt through execution, providing complete runtime visibility and an auditable record of agent behavior."Source: [hiddenlayer.com](https://www.hiddenlayer.com/solutions/agent-harness-security), read 2026-11-16. - **hiddenlayer-30** (observed): HiddenLayer publishes product documentation, an index and release notes openly at docs.hiddenlayer.ai without sign-in.Quote: "Fetch the complete documentation index at:"Source: [docs.hiddenlayer.ai](https://docs.hiddenlayer.ai/docs/products/runtime/releasenotes.md), read 2026-11-16. - **hiddenlayer-31** (none-found): None found on 2026-11-16: a published price or pricing page; the /pricing URL returned Not Found and the only sales path found is a demo request.Source: [hiddenlayer.com](https://www.hiddenlayer.com/pricing), read 2026-11-16. - **hiddenlayer-32** (vendor): HiddenLayer publishes an annual AI Threat Landscape Report with survey data.Quote: "Get your copy of the 2026 AI Threat Landscape Report for survey data, technical research, and case studies behind these findings."Source: [hiddenlayer.com](https://www.hiddenlayer.com/report-and-guide/threatreport2026), read 2026-11-16. - **hiddenlayer-33** (observed): HiddenLayer publishes Apache-2.0 licensed integration code on GitHub, such as guardrails for the OpenAI Agents SDK, which sends agent input, output and every tool call to its analyze endpoint.Quote: "Agent input/output and every tool call are sent through HiddenLayer's analyze endpoint so prompt-injection and policy violations are caught automatically."Source: [raw.githubusercontent.com](https://raw.githubusercontent.com/hiddenlayerai/hiddenlayer-openai-guardrails/main/README.md), read 2026-11-16. - **hiddenlayer-34** (vendor): On 2026-09-02, HiddenLayer announced a $100 million Series B led by Delta-v Capital, with Ten Eleven Ventures, Morgan Stanley, M12 and Booz Allen Ventures participating.Quote: "today announced a $100 million Series B funding round led by"Source: [hiddenlayer.com](https://www.hiddenlayer.com/news/hiddenlayer-100m-series-b-ai-security), read 2026-11-16. - **hiddenlayer-35** (vendor): On 2026-09-02, HiddenLayer said ARR grew more than 10x and it signed more than 50 new platform customers, including banks, insurers, government and US defense and intelligence communities.Quote: "signed more than 50 new platform customers"Source: [hiddenlayer.com](https://www.hiddenlayer.com/news/hiddenlayer-100m-series-b-ai-security), read 2026-11-16. - **hiddenlayer-36** (third-party): On 2026-09-02, TechCrunch reported that HiddenLayer raised a $100 million Series B led by Delta-v Capital, and that its CEO likened runtime security to EDR for AI.Quote: "the startup has now raised $100 million in a Series B funding round that was led by Delta-v Capital"Source: [techcrunch.com](https://techcrunch.com/2026/09/02/hiddenlayer-nabs-100m-as-enterprises-rush-to-secure-their-ai-deployments/), read 2026-11-16. - **hiddenlayer-37** (third-party): TechCrunch reported that HiddenLayer has contracts with the Department of Defense and intelligence community, and names no individual customer; customer logos on the product page are quotes from CISOs at NFL, GitLab and others.Quote: "it also has contracts with the Department of Defense and intelligence community"Source: [techcrunch.com](https://techcrunch.com/2026/09/02/hiddenlayer-nabs-100m-as-enterprises-rush-to-secure-their-ai-deployments/), read 2026-11-16. - **hiddenlayer-38** (vendor): HiddenLayer's runtime page carries named executive endorsements, including Tomas Maldonado, CISO, NFL and Josh Lemos, CISO, GitLab.Quote: "Tomas Maldonado"Source: [hiddenlayer.com](https://www.hiddenlayer.com/platform/ai-runtime-security), read 2026-11-16. - **hiddenlayer-39** (none-found): None found on 2026-11-16: independent analyst coverage (for example a Gartner or Forrester report page) of HiddenLayer on the pages read.Source: [hiddenlayer.com](https://www.hiddenlayer.com/newsroom), read 2026-11-16. - **hiddenlayer-40** (vendor): HiddenLayer is a founding member of the ServiceNow Control Tower Alliance, announced by ServiceNow in a Business Wire release reposted on the HiddenLayer newsroom.Quote: "HiddenLayer joins ServiceNow’s Control Tower Alliance as a founding member"Source: [hiddenlayer.com](https://www.hiddenlayer.com/newsroom), read 2026-11-16. - **hiddenlayer-41** (third-party): On 2026-09-03, FinTech Global reported that HiddenLayer's Agent Harness Security builds on the Runtime Security module to protect AI coding agents while they operate.Quote: "extending protection to AI coding agents while they operate"Source: [fintech.global](https://fintech.global/2026/09/03/hiddenlayer-bags-100m-to-guard-enterprise-ai-agents/), read 2026-11-16. ## News mentioning HiddenLayer - [HiddenLayer raises 00M Series B for agent runtime](/news/hiddenlayer-100m-series-b/) ## Related pages - [All vendor profiles](/vendors/) - [CrowdStrike profile](/vendors/crowdstrike/) - [Lasso Security profile](/vendors/lasso-security/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # Lasso Security AI runtime security URL: https://airuntimesecuritytools.com/vendors/lasso-security/ Updated: 2026-11-16 Lasso publishes numbers where most vendors do not. It says gateway enforcement, which blocks, masks or alerts, completes in under 50 ms, and SiliconANGLE reported on 2 Sep 2026 a CPU-only guardrail returning a verdict in under 5 ms, per Lasso's own benchmarks. Coverage spans model requests and responses, gateway-fronted applications, agent tool calls and on-device hooks for Claude Code. Lasso says it starts in hours with one line of code, and it streams audit trails to a SIEM. Gaps: fail-open or fail-closed behavior, package install blocking, network call blocking, public documentation and a price are all none found on 16 Nov 2026. ## Strength and weakness for AI Runtime Security Tools **Strength:** Publishes inline block, mask or alert enforcement with under 50 ms and a sub-5 ms CPU guardrail claim (lasso-security-7, lasso-security-6, lasso-security-14). **Weakness:** No fail-open or fail-closed behavior published and package install and network call blocking not named (lasso-security-9; none found on 2026-11-16). ## Scores and evidence Lasso Security is 3rd in the overall ranking at 63.5 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Lasso Security: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Action types stopped before they complete | 7 of 10 | 30 | The vendor claims blocking of out-of-scope tool calls, prompt injection, sensitive uploads, credential leakage in tool calls and responses before they return. All of it is vendor-only. | [lasso-security-12](#fact-lasso-security-12), [lasso-security-18](#fact-lasso-security-18), [lasso-security-21](#fact-lasso-security-21), [lasso-security-17](#fact-lasso-security-17), [lasso-security-23](#fact-lasso-security-23), [lasso-security-14](#fact-lasso-security-14), [lasso-security-10](#fact-lasso-security-10) | | Layers covered | 7 of 10 | 20 | The vendor claims model request and response inspection, gateway coverage for applications, agent tool calls and on-device hooks. All of it is vendor-only. | [lasso-security-15](#fact-lasso-security-15), [lasso-security-16](#fact-lasso-security-16), [lasso-security-12](#fact-lasso-security-12), [lasso-security-28](#fact-lasso-security-28), [lasso-security-29](#fact-lasso-security-29) | | Latency and failure mode | 5 of 10 | 15 | The vendor publishes under 50 ms gateway enforcement and third-party press reports under 5 ms per the vendor's benchmarks. Fail-open or fail-closed behavior: none found. | [lasso-security-07](#fact-lasso-security-07), [lasso-security-06](#fact-lasso-security-06), [lasso-security-09](#fact-lasso-security-09) | | Deployment | 7 of 10 | 15 | The vendor says it starts in hours with one line of code via gateway, API or SDK, plus hooks through enterprise management. All of it is vendor-claimed. | [lasso-security-47](#fact-lasso-security-47), [lasso-security-48](#fact-lasso-security-48), [lasso-security-46](#fact-lasso-security-46), [lasso-security-16](#fact-lasso-security-16), [lasso-security-29](#fact-lasso-security-29) | | Detection and response | 7 of 10 | 5 | The vendor claims audit trails, SIEM streaming, alerts with trace context and an execution graph. SOAR or ticketing: none found. | [lasso-security-25](#fact-lasso-security-25), [lasso-security-41](#fact-lasso-security-41), [lasso-security-11](#fact-lasso-security-11), [lasso-security-38](#fact-lasso-security-38), [lasso-security-40](#fact-lasso-security-40), [lasso-security-42](#fact-lasso-security-42) | | Transparency | 3 of 10 | 10 | Research is published. Public documentation and a published price: none found. | [lasso-security-49](#fact-lasso-security-49), [lasso-security-50](#fact-lasso-security-50), [lasso-security-51](#fact-lasso-security-51), [lasso-security-52](#fact-lasso-security-52) | | Proof | 8 of 10 | 5 | Third-party press confirms funding and names customers, and analyst recognition is vendor-reported. | [lasso-security-02](#fact-lasso-security-02), [lasso-security-03](#fact-lasso-security-03), [lasso-security-04](#fact-lasso-security-04) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 3rd | 63.5 | | [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) | 4th | 66.5 | | [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) | joint 3rd | 63.5 | | [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) | 3rd | 67.5 | | [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) | joint 3rd | 66.5 | | [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) | 4th | 66.5 | ## Product, deployment, price and owner - Product assessed: Lasso AI Detection & Response with Intent Security Engine and LEAP/RAPID guardrails - Deployment: Agentless (read-only CI and cloud integrations), API, SDK, gateway or firewall integration, browser extension, and Claude Code lifecycle hooks via enterprise management - Time to deploy: Lasso says start in hours, not weeks, and usage control in minutes - Price: none found on 2026-11-16 - Owner: Lasso Security Ltd. - Website: [lasso.security](https://www.lasso.security) ## Facts and sources - **lasso-security-01** (vendor): Lasso says its AI Security Platform gives enterprises visibility, control and protection across AI models, agents and apps.Quote: "Lasso's AI Security Platform gives enterprises visibility, control, and protection across AI models, agents, and apps."Source: [lasso.security](https://www.lasso.security/), read 2026-11-16. - **lasso-security-02** (third-party): SiliconANGLE reports that Lasso Security Ltd. was founded in 2023 by Schulman and Dror and has raised more than $37 million in total, including a $30 million round announced on 2026-09-02 led by ClearSky Advisors.Quote: "Investors have now put more than $37 million into the company."Source: [siliconangle.com](https://siliconangle.com/2026/09/02/lasso-security-launches-cpu-only-ai-guardrail-raises-30m/), read 2026-11-16. - **lasso-security-03** (third-party): SiliconANGLE names BMW, eToro, Fiverr and the U.S. Department of Homeland Security as Lasso customers (reported 2026-09-02).Quote: "BMW AG is a customer, and so are eToro Group Ltd., Fiverr International Ltd. and the U.S. Department of Homeland Security."Source: [siliconangle.com](https://siliconangle.com/2026/09/02/lasso-security-launches-cpu-only-ai-guardrail-raises-30m/), read 2026-11-16. - **lasso-security-04** (vendor): On 2026-09-15 Lasso said Latio Tech's 2026 AI Security Market Report named it a platform leader (vendor's own account of a third-party report).Quote: "In the report, Lasso is named as one of the major Platform Leaders."Source: [lasso.security](https://www.lasso.security/blog/lasso-named-a-leader-in-latios-2026-ai-security-market-report), read 2026-11-16. - **lasso-security-05** (third-party): On 2026-09-01 TechCentral (a promoted article by NEWORDER) said NEWORDER is Lasso's official partner in South Africa and that Lasso is certified to ISO 27001 and SOC 2.Quote: "Lasso Security was founded in Tel Aviv in 2023, is certified to ISO 27001 and SOC 2, and has customers across Europe, the US and Israel."Source: [techcentral.co.za](https://techcentral.co.za/neworder-lasso-security-ai-agent-controls/285558/), read 2026-11-16. - **lasso-security-06** (third-party): On 2026-09-02 SiliconANGLE reported Lasso's new LEAP guardrail returns a verdict in under five milliseconds on CPUs, per Lasso's own benchmarks.Quote: "LEAP is transformer-free and returns a verdict in under five milliseconds, the company said, at thousands of times the throughput of existing guardrails."Source: [siliconangle.com](https://siliconangle.com/2026/09/02/lasso-security-launches-cpu-only-ai-guardrail-raises-30m/), read 2026-11-16. - **lasso-security-07** (vendor): Lasso says its gateway enforcement (block, mask or alert) completes in under 50 ms.Quote: "All enforcement executes in under 50ms, with no manual intervention required."Source: [lasso.security](https://www.lasso.security/use-cases/ai-gateway-security), read 2026-11-16. - **lasso-security-08** (vendor): Lasso says its LLM-as-judge enforcement runs 570 times faster than standard LLM inference, with no latency impact on the user.Quote: "Lasso’s LLM as a judge operates at 570x the speed of standard LLM inference, so enforcement is contextual and intent-aware with no latency impact on the user experience."Source: [lasso.security](https://www.lasso.security/platform/ai-detection-response), read 2026-11-16. - **lasso-security-09** (none-found): None found on 2026-11-16: fail-open or fail-closed behavior, or timeout behavior when the control fails, searched on the AI Detection & Response page and the gateway, LEAP and endpoint pages.Source: [lasso.security](https://www.lasso.security/platform/ai-detection-response), read 2026-11-16. - **lasso-security-10** (vendor): Lasso says it enforces policy inline so a harmful agent action can be stopped before it executes, and that it can block or quarantine an action.Quote: "preventing unauthorized actions requires inline enforcement that can stop a harmful action before it executes and limit what each agent is permitted to do."Source: [lasso.security](https://www.lasso.security/platform/ai-detection-response), read 2026-11-16. - **lasso-security-11** (vendor): Lasso says on detection it can block or quarantine an action before it executes, alert with trace context, and feed findings to SIEM.Quote: "Lasso can block or quarantine an action before it executes, alert security teams with full trace context for investigation, and feed findings into your existing SIEM and observability workflows."Source: [lasso.security](https://www.lasso.security/platform/ai-detection-response), read 2026-11-16. - **lasso-security-12** (vendor): Lasso says it inspects every coding-agent tool call before execution and flags or blocks anything outside the permitted scope.Quote: "Lasso inspects every tool call before it is executed."Source: [lasso.security](https://www.lasso.security/use-cases/ai-coding-assistants), read 2026-11-16. - **lasso-security-13** (vendor): Lasso says its coding-assistant coverage addresses agents that read codebases, execute shell commands and call MCP servers (Claude Code, Cursor, Codex, OpenCode).Quote: "Claude Code, Cursor, and Codex read your codebase, executes shell commands, calls MCPs, and takes actions on your behalf."Source: [lasso.security](https://www.lasso.security/use-cases/ai-coding-assistants), read 2026-11-16. - **lasso-security-14** (vendor): Lasso says gateway mode can block an interaction before it reaches the model or before a response returns, mask sensitive data in transit, or allow and alert.Quote: "block the interaction before it reaches the model or before a response is returned, mask sensitive data in transit, or allow and alert where interruption would affect a critical workflow"Source: [lasso.security](https://www.lasso.security/use-cases/ai-gateway-security), read 2026-11-16. - **lasso-security-15** (vendor): Lasso says it inspects every request before it reaches the model and every response before it returns.Quote: "Inspect every request inline before it reaches the model and every response before it returns."Source: [lasso.security](https://www.lasso.security/use-cases/ai-gateway-security), read 2026-11-16. - **lasso-security-16** (vendor): Lasso says it integrates with Kong, Envoy AI Gateway, TrueFoundry, Portkey, LiteLLM and other AI gateways without source code changes.Quote: "Lasso integrates with Kong, Envoy AI Gateway, TrueFoundry, Portkey, LiteLLM, and others."Source: [lasso.security](https://www.lasso.security/use-cases/ai-gateway-security), read 2026-11-16. - **lasso-security-17** (vendor): Lasso says it detects sensitive data (PII, PCI) in prompts and file uploads and masks it before the prompt reaches third-party models.Quote: "Detect and monitor in real time if PII, PCI, or any other sensitive data is shared in prompts or file uploads, and mask those contents before the prompt is sent"Source: [lasso.security](https://www.lasso.security/platform/ai-usage-control), read 2026-11-16. - **lasso-security-18** (vendor): Lasso says it blocks injected or guardrail-bypassing inputs and decodes over 3,000 obfuscation techniques at the intent layer.Quote: "Lasso decodes over 3,000 obfuscation techniques and detects threats at the intent layer. It blocks inputs designed to bypass safety guardrails or trick agents into performing unauthorized actions"Source: [lasso.security](https://www.lasso.security/platform/ai-agents-security), read 2026-11-16. - **lasso-security-21** (vendor): Lasso says it scans MCP tool descriptions for hidden instructions and blocks prompt injection at the connection layer.Quote: "Scan MCP server tool descriptions for hidden instructions. Detect and block prompt injection attempts at the connection layer."Source: [lasso.security](https://www.lasso.security/use-cases/mcp-security), read 2026-11-16. - **lasso-security-22** (vendor): Lasso says it monitors every MCP tool call in real time and can mask sensitive data before it reaches external servers.Quote: "Monitor every MCP tool call in real-time to identify indirect prompt injection, memory poisoning, and other AI threats or attack techniques."Source: [lasso.security](https://www.lasso.security/use-cases/mcp-security), read 2026-11-16. - **lasso-security-23** (vendor): Lasso says it detects PII, API keys and credentials shared through MCP tool calls and masks them before they reach external servers.Quote: "Detect and monitor in real-time if PII, API keys, credentials, or any other sensitive data is shared through MCP tool calls. Mask sensitive contents before they reach external servers."Source: [lasso.security](https://www.lasso.security/use-cases/mcp-security), read 2026-11-16. - **lasso-security-25** (vendor): Lasso says it keeps audit trails of every MCP connection and tool call and exports logs to a SIEM.Quote: "Maintain complete audit trails of every MCP connection and tool call. Export logs to your SIEM."Source: [lasso.security](https://www.lasso.security/use-cases/mcp-security), read 2026-11-16. - **lasso-security-26** (vendor): Lasso says it governs AI activity on employee endpoints agentlessly, so visibility into AI tools, coding agents and MCP connections starts on day one.Quote: "Lasso governs AI activity on employee endpoints through an agentless approach, so security teams get visibility into AI tools, coding agents, and MCP connections from day one"Source: [lasso.security](https://www.lasso.security/blog/securing-ai-on-endpoints-how-to-govern-ai-tools-coding-agents-and-mcp-connections), read 2026-11-16. - **lasso-security-28** (vendor): Lasso says policy enforcement happens at the execution layer on the device itself.Quote: "Policy enforcement happens at the execution layer on the device itself, so protection holds even when an agent's activity would not otherwise pass through a monitored network path."Source: [lasso.security](https://www.lasso.security/blog/securing-ai-on-endpoints-how-to-govern-ai-tools-coding-agents-and-mcp-connections), read 2026-11-16. - **lasso-security-29** (vendor): Lasso says it attaches to Claude Code, Cursor and Codex through lifecycle hooks deployed by the enterprise management platform, with no agent to install.Quote: "Lasso deploys through the enterprise management platform, so there is no agent to install and no code to change."Source: [lasso.security](https://www.lasso.security/use-cases/ai-coding-assistants), read 2026-11-16. - **lasso-security-30** (none-found): None found on 2026-11-16: supported operating systems and a time to deploy for MCP discovery on developer machines, searched on the MCP security, coding assistants and endpoint pages.Source: [lasso.security](https://www.lasso.security/use-cases/mcp-security), read 2026-11-16. - **lasso-security-33** (vendor): Lasso says its discovery is agentless and read-only, using CI pipelines and native cloud and agent-builder integrations.Quote: "Agentless, read-only integration into CI pipelines and native integrations with cloud providers and third-party agent builder tools."Source: [lasso.security](https://www.lasso.security/platform/discovery-ai-bom), read 2026-11-16. - **lasso-security-36** (vendor): Lasso says AI usage control deploys through a browser extension, proxy integrations and data integrations, and can be running in minutes.Quote: "Enable full AI usage control in minutes through Lasso’s lightweight deployment paths like our browser extension, proxy integrations, and data integrations."Source: [lasso.security](https://www.lasso.security/platform/ai-usage-control), read 2026-11-16. - **lasso-security-38** (vendor): Lasso says it maps tool calls, RAG retrievals and sub-agent communications into an execution graph.Quote: "Lasso monitors every interaction across your AI stack, from tool calls and RAG retrievals to sub-agent communications."Source: [lasso.security](https://www.lasso.security/platform/ai-detection-response), read 2026-11-16. - **lasso-security-40** (vendor): Lasso says exportable audit trails are updated on every interaction and mapped to NIST AI RMF, OWASP, EU AI Act, ISO 42001 and others.Quote: "Exportable audit trails are updated on every interaction, not generated on demand when an audit request arrives."Source: [lasso.security](https://www.lasso.security/use-cases/ai-gateway-security), read 2026-11-16. - **lasso-security-41** (vendor): Lasso says alerts and audit logs can be viewed in Lasso or streamed to a SIEM.Quote: "View alerts and audit logs directly in Lasso, or stream findings into your SIEM to investigate agent incidents alongside the rest of your security telemetry."Source: [lasso.security](https://www.lasso.security/platform/ai-detection-response), read 2026-11-16. - **lasso-security-42** (none-found): None found on 2026-11-16: documented ticketing or identity-provider integrations, or SOAR export, searched on the detection and response, usage control and MCP pages.Source: [lasso.security](https://www.lasso.security/platform/ai-detection-response), read 2026-11-16. - **lasso-security-46** (vendor): Lasso says it deploys via API, firewall or AI gateway integration with no architectural changes.Quote: "Deploy via API, Firewall, or AI Gateway Integration"Source: [lasso.security](https://www.lasso.security/platform/ai-detection-response), read 2026-11-16. - **lasso-security-47** (vendor): Lasso says it runs from the Lasso tenant and starts in hours, not weeks, with no agents or code changes.Quote: "No agents. No code changes. No source code access. Run from the Lasso tenant and start in hours, not weeks."Source: [lasso.security](https://www.lasso.security/platform/ai-detection-response), read 2026-11-16. - **lasso-security-48** (vendor): Lasso says it deploys with a single line of code via Gateway, API or SDK and ships policy templates aligned to OWASP, MITRE and NIST.Quote: "Lasso deploys with a single line of code via Gateway, API, or SDK."Source: [lasso.security](https://www.lasso.security/platform/ai-agents-security), read 2026-11-16. - **lasso-security-49** (vendor): On 2026-09-17 Lasso published research on how LLM text watermarking changes refusals and agent tool calling.Quote: "We find that it does, in both model refusal behavior and agent tool calling."Source: [lasso.security](https://www.lasso.security/blog/the-provenance-tax-understanding-the-impact-of-llm-watermarking-on-ai-agent-behavior), read 2026-11-16. - **lasso-security-50** (vendor): On 2026-09-09 Lasso published a technical post on LEAP, its CPU-based guardrail, with benchmark claims.Quote: "Introducing LEAP: CPU-based AI Security Guardrails with GPU-Class Accuracy"Source: [lasso.security](https://www.lasso.security/blog/introducing-leap-cpu-based-ai-security-guardrails-with-gpu-class-accuracy), read 2026-11-16. - **lasso-security-51** (none-found): None found on 2026-11-16: a published price; the pricing URL returns a 404 page and the site routes to Book a Demo.Source: [lasso.security](https://www.lasso.security/pricing), read 2026-11-16. - **lasso-security-52** (none-found): None found on 2026-11-16: public product documentation; docs.lasso.security loads only a sign-in dashboard titled Lasso Security - Dashboard.Source: [docs.lasso.security](https://docs.lasso.security), read 2026-11-16. - **lasso-security-53** (vendor): Lasso says its detection covers indirect prompt injection in files, web responses and MCP outputs before they reach the coding agent.Quote: "detect injected payloads or hidden instructions in files, web responses, and MCP outputs before they reach the coding agent’s decision layer"Source: [lasso.security](https://www.lasso.security/use-cases/ai-coding-assistants), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [HiddenLayer profile](/vendors/hiddenlayer/) - [Microsoft profile](/vendors/microsoft/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # Microsoft AI runtime security URL: https://airuntimesecuritytools.com/vendors/microsoft/ Updated: 2026-11-16 Microsoft documents controls at several layers, and several are still in preview. Prompt Shields in Azure AI Content Safety detect user prompt and document attacks, and in Foundry can block flagged inputs before the model processes them. Defender for Endpoint runtime protection inspects the agent loop and can block risky activity before it runs, but it is public preview, with listed limits for certificate pinning, HTTP/3 and Claude Code. Alerts flow into Defender XDR. Agent 365 is listed at 15 dollars per user per month. Gaps from 16 Nov 2026: a latency figure in milliseconds, fail-open or fail-closed behavior and generally available endpoint blocking are none found. ## Strength and weakness for AI Runtime Security Tools **Strength:** Defender can block prompt injection and risky agent actions before they run on the endpoint, and Prompt Shields block attack inputs in Foundry (microsoft-19, microsoft-41). **Weakness:** None found on 2026-11-16 for published latency in milliseconds or fail-open and fail-closed behavior; endpoint blocking is preview with a documented Claude Code limit (microsoft-46, microsoft-21, microsoft-23). ## Scores and evidence Microsoft is 15th in the overall ranking at 45.0 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Microsoft: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Action types stopped before they complete | 4 of 10 | 30 | Blocking of prompt attacks in Foundry and policy evaluation of tool invocations are documented; endpoint runtime blocking is public preview and the reported runtime stop has no availability date, so both count only as mentions. Blocking of shell, file, package and network actions as generally available: none found. | [microsoft-41](#fact-microsoft-41), [microsoft-30](#fact-microsoft-30), [microsoft-19](#fact-microsoft-19), [microsoft-26](#fact-microsoft-26), [microsoft-63](#fact-microsoft-63), [microsoft-66](#fact-microsoft-66), [microsoft-10](#fact-microsoft-10) | | Layers covered | 6 of 10 | 20 | Model, application (API, where the customer application decides) and agent tool-invocation layers are documented. Endpoint enforcement is public preview only, so generally available endpoint enforcement: none found. | [microsoft-41](#fact-microsoft-41), [microsoft-30](#fact-microsoft-30), [microsoft-45](#fact-microsoft-45), [microsoft-42](#fact-microsoft-42), [microsoft-19](#fact-microsoft-19) | | Latency and failure mode | 0 of 10 | 15 | The vendor says latency is minimal, but a millisecond figure and fail-open or fail-closed behavior: none found. | [microsoft-21](#fact-microsoft-21), [microsoft-46](#fact-microsoft-46) | | Deployment | 6 of 10 | 15 | Documented operating systems, Intune or PowerShell onboarding and API delivery. Published time to deploy: none found. | [microsoft-13](#fact-microsoft-13), [microsoft-25](#fact-microsoft-25), [microsoft-45](#fact-microsoft-45), [microsoft-48](#fact-microsoft-48) | | Detection and response | 6 of 10 | 5 | Alerts flow into the vendor's own XDR incidents and agent data is queryable in advanced hunting. The block record and audit mode belong to the preview runtime protection, and export to an outside SIEM or SOAR: none found. | [microsoft-47](#fact-microsoft-47), [microsoft-17](#fact-microsoft-17), [microsoft-24](#fact-microsoft-24), [microsoft-22](#fact-microsoft-22) | | Transparency | 6 of 10 | 10 | Documentation is public and some prices are published, and the MCP gateway source is observed on GitHub. Published research: none found, and third-party press reports no pricing for the new agent controls. | [microsoft-52](#fact-microsoft-52), [microsoft-49](#fact-microsoft-49), [microsoft-50](#fact-microsoft-50), [microsoft-53](#fact-microsoft-53), [microsoft-66](#fact-microsoft-66) | | Proof | 6 of 10 | 5 | A large public company with third-party coverage. Named customers or analyst coverage for these AI runtime products: none found. | [microsoft-64](#fact-microsoft-64), [microsoft-58](#fact-microsoft-58) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 15th | 45.0 | | [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) | 15th | 47.0 | | [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) | 15th | 40.0 | | [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) | joint 14th | 51.0 | | [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) | 14th | 49.0 | | [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) | 15th | 49.0 | ## Product, deployment, price and owner - Product assessed: Defender for Endpoint AI agent runtime protection (preview), Azure AI Content Safety Prompt Shields, Defender for Cloud AI threat protection, Agent 365 real-time protection - Deployment: Cloud-managed Defender for Endpoint client onboarding (Intune or PowerShell for runtime protection); API-based Content Safety; Purview browser extension; Entra and Defender portal controls - Time to deploy: none found on 2026-11-16 (discovery starts automatically after device onboarding; Purview policies need at least 24 hours) - Price: Agent 365 is listed at $15.00 per user per month, paid yearly; Content Safety Standard shows a dash; Defender for Endpoint and Cloud Apps prices none found on 2026-11-16 on readable pages. - Owner: Microsoft Corporation - Website: learn.microsoft.com ## Facts and sources - **microsoft-10** (vendor): Microsoft documents Endpoint DLP detection of pasting content to restricted service domains in supported browsers, evaluated on the pasted content.Quote: "Detects when a user attempts to paste content to a restricted service domain. Evaluation is performed on the content that is being pasted."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/purview/endpoint-dlp-learn-about), read 2026-11-16. - **microsoft-12** (vendor): Microsoft says local agent discovery needs no extra deployment beyond onboarding to Defender for Endpoint and starts automatically; Plan 2 is the minimum license.Quote: "You don't need additional deployment, configuration, or scripts beyond the device onboarding requirements. If the device meets all of the prerequisites, agent discovery begins automatically."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/discover-local-ai-agents), read 2026-11-16. - **microsoft-13** (vendor): Microsoft lists Windows, macOS, Linux, Android and iOS as Defender for Endpoint operating systems; local agent discovery is documented for Windows and macOS.Quote: "Microsoft Defender for Endpoint supports the following operating systems: Windows, macOS, Linux, Android, and iOS."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/microsoft-defender-endpoint), read 2026-11-16. - **microsoft-17** (vendor): Microsoft says local agent data can be queried in advanced hunting, including an AgentsInfo table that carries MCP servers and declared tools.Quote: "Use advanced hunting to proactively investigate local AI agent presence, understand how agents are configured, and identify the agents and users that carry the most risk."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/discover-local-ai-agents), read 2026-11-16. - **microsoft-19** (vendor): Microsoft says Defender for Endpoint runtime protection inspects the agent loop on endpoints and can block risky activity before it executes; this is a public preview feature.Quote: "Microsoft Defender inspects the agent loop (user prompts, tool calls, and tool responses) and can block risky activity before it executes"Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/whats-new-in-microsoft-defender-endpoint), read 2026-11-16. - **microsoft-20** (vendor): Microsoft documents agent-native event inspection for agents that expose vendor event interfaces (Claude Code, Codex CLI, GitHub Copilot CLI) and network inspection for others.Quote: "Agents such as Claude Code, Codex CLI, and GitHub Copilot CLI expose these interfaces, and Defender uses them to inspect agent activity"Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-runtime-protection-overview), read 2026-11-16. - **microsoft-21** (vendor): Microsoft says each runtime scan is an inline check with minimal added latency, but publishes no millisecond figure on this page.Quote: "Each scan is a fast, inline check rather than continuous monitoring of the agent process, so the added latency is minimal."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-runtime-protection-overview), read 2026-11-16. - **microsoft-22** (vendor): Microsoft documents block, audit and disabled modes, advises starting in audit, and protects the setting with tamper protection.Quote: "Microsoft recommends starting in audit mode to observe detections and validate accuracy before switching to block mode for active enforcement."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-runtime-protection-overview), read 2026-11-16. - **microsoft-23** (vendor): Microsoft lists preview limits: network inspection skips agents using certificate pinning or HTTP/3, and Claude Code may keep processing a tool response in Block mode after a post-tool-response block.Quote: "Network inspection doesn't support agents that use certificate pinning or HTTP/3."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-runtime-protection-overview), read 2026-11-16. - **microsoft-24** (vendor): Microsoft says a block is notified to the user in the agent UI and by Windows toast, recorded in protection history, and sent as a security alert correlated into incidents.Quote: "a security alert is sent to Defender, correlated into incidents for the SOC to investigate."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-runtime-protection-overview), read 2026-11-16. - **microsoft-25** (vendor): Microsoft says runtime protection is configured in audit or block mode through local PowerShell or organization-wide Intune.Quote: "For local PowerShell and organization-wide Microsoft Intune configuration options"Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-runtime-protection-overview), read 2026-11-16. - **microsoft-26** (vendor): The support matrix lists runtime protection per agent on Windows (preview), including Claude Code, GitHub Copilot CLI and app, Claude Desktop Code, VS Code Copilot and Claude Code extensions, plus OpenClaw, Ollama Desktop and ChatGPT Desktop via network inspection.Quote: "Network inspection covers agents that communicate with LLMs through well-known inference services. It doesn't cover agents that use certificate pinning or HTTP/3."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-support-matrix), read 2026-11-16. - **microsoft-28** (vendor): Microsoft Defender integrates with Agent 365 so that real-time protection evaluates tool calls and responses for Work IQ MCP and customer MCP tools onboarded to Agent 365; the default rule audits and custom rules can block before execution.Quote: "custom rules can block matching actions before they execute"Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-cloud-apps/ai-agent-protection), read 2026-11-16. - **microsoft-30** (vendor): Microsoft says real-time protection evaluates tool invocations against policy and near-real-time detections cover indirect prompt injection, secret leakage and suspicious IP access.Quote: "Near-real-time detections analyze Agent 365 observability data for threats such as indirect prompt injection, evasion techniques, malicious content propagation, secret leakage"Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-cloud-apps/ai-agent-protection), read 2026-11-16. - **microsoft-41** (vendor): Microsoft says Prompt Shields detect user prompt attacks and document attacks, and in Foundry, with blocking enabled, block inputs classified as attacks before the model processes them.Quote: "When you enable blocking in Foundry, Prompt Shields block inputs classified as attacks before those inputs are processed by the model."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection), read 2026-11-16. - **microsoft-42** (vendor): Microsoft says in the Azure AI Content Safety API the service does not itself block or route content; the customer application decides.Quote: "In Azure AI Content Safety APIs, the API doesn't itself block or route content."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection), read 2026-11-16. - **microsoft-43** (vendor): Microsoft says Prompt Shields can scan document attacks at the tool response point for agents on supported tools (Preview) and that document attack detection is off by default.Quote: "For agents, Prompt Shields can also scan document attacks at the tool response intervention point for supported tools (Preview)."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection), read 2026-11-16. - **microsoft-44** (vendor): Microsoft states Prompt Shields do not judge whether content is harmful or accurate and that performance is primarily trained on English.Quote: "the user prompt attack and document attack models are primarily trained and evaluated on English content."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection), read 2026-11-16. - **microsoft-45** (vendor): Microsoft says guardrails work inside Microsoft Foundry or as a standalone offering through Content Safety Studio and APIs for models and agents outside Foundry.Quote: "Your models and agents are outside Foundry - or span multiple platforms - and you want a consistent guardrail experience."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/overview), read 2026-11-16. - **microsoft-46** (none-found): None found on 2026-11-16: no published latency figure or fail-open or fail-closed behavior for Prompt Shields or Content Safety on the Prompt Shields and overview pages.Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection), read 2026-11-16. - **microsoft-47** (vendor): Microsoft says Defender for Cloud AI threat protection works with Prompt Shields to alert on data leakage, jailbreak and credential theft, and sends alerts to Defender XDR; it supports text tokens only.Quote: "provide security alerts for threats like data leakage, data poisoning, jailbreak, credential theft, and more"Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/azure/defender-for-cloud/ai-threat-protection), read 2026-11-16. - **microsoft-48** (vendor): Microsoft says Defender for AI Services has a 30-day free trial capped at 75 billion tokens scanned and is limited to commercial clouds.Quote: "Defender for AI Services includes a 30-day free trial, capped at 75 billion tokens scanned."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/azure/defender-for-cloud/ai-threat-protection), read 2026-11-16. - **microsoft-49** (vendor): Microsoft lists Agent 365 at $15.00 per user per month, paid yearly, on its product page.Quote: "$15.00 user/month, paid yearly (Annual commitment)"Source: [microsoft.com](https://www.microsoft.com/en-us/microsoft-agent-365), read 2026-11-16. - **microsoft-50** (vendor): Microsoft's Azure AI Content Safety pricing page shows a free tier of 5,000 text records per month and a Standard tier whose price displays as a dash.The exact quote is on the source page.Source: [azure.microsoft.com](https://azure.microsoft.com/en-us/pricing/details/content-safety/), read 2026-11-16. - **microsoft-51** (none-found): None found on 2026-11-16: no price for Defender for Cloud Apps or Defender for Endpoint standalone is readable on the learn pages; they point to Microsoft 365 E5 inclusion and separate pricing pages that returned 403 to a script.Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/microsoft-defender-endpoint), read 2026-11-16. - **microsoft-52** (vendor): Microsoft documentation on learn.microsoft.com is readable without sign-in, with last-updated dates (Agent 365 overview 2026-11-12).Quote: "Last updated on 2026-11-12"Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/microsoft-agent-365/overview), read 2026-11-16. - **microsoft-53** (observed): Microsoft's MCP Gateway source is public on GitHub under the MIT license and offers telemetry, access control and observability hooks.Quote: "Enterprise-ready integration points including telemetry, access control and observability."Source: [github.com](https://github.com/microsoft/mcp-gateway), read 2026-11-16. - **microsoft-56** (third-party): On 2026-10-20, Help Net Security quoted Satya Nadella, CEO of Microsoft, on treating frontier models as insider risks, and said Project Perception shares signals with Agent 365.Quote: "Satya Nadella , CEO of Microsoft, wrote"Source: [helpnetsecurity.com](https://www.helpnetsecurity.com/2026/10/20/microsofts-ai-agents-threats-guardian-agents/), read 2026-11-16. - **microsoft-58** (none-found): None found on 2026-11-16: no named customer list or analyst page for shadow AI discovery, MCP security or AI runtime security was read on the Microsoft pages fetched.Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/whats-new-in-microsoft-defender-endpoint), read 2026-11-16. - **microsoft-59** (vendor): Microsoft owns and sells these products itself: Defender for Endpoint is described as the endpoint security pillar of Microsoft Defender.Quote: "As the endpoint security pillar of Microsoft Defender , Defender for Endpoint feeds endpoint signals into the unified Defender portal."Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/defender-endpoint/microsoft-defender-endpoint), read 2026-11-16. - **microsoft-61** (vendor): Microsoft Purview says DSPM for AI policies need at least 24 hours to collect data before results appear.Quote: "Allow at least 24 hours for these new policies to collect data to display the results in Data Security Posture Management for AI"Source: [learn.microsoft.com](https://learn.microsoft.com/en-us/purview/dspm-for-ai), read 2026-11-16. - **microsoft-63** (third-party): On 2026-10-19, SiliconANGLE reported Microsoft extended Conditional Access to agents so that rising risk can hold higher-impact actions for human approval and, at the top of the scale, isolate or stop the runtime.Quote: "Higher-impact actions can be held for verified human approval, and at the top of the scale the runtime is isolated or stopped entirely."Source: [siliconangle.com](https://siliconangle.com/2026/10/19/microsoft-targets-the-agentic-insider-with-new-controls-for-rogue-ai-agents/), read 2026-11-16. - **microsoft-64** (third-party): On 2026-10-19, SiliconANGLE covered the launch and named Satya Nadella as Microsoft Chairman and Chief Executive and Hayete Gallot as executive vice president of Microsoft Security.Quote: "Microsoft Chairman and Chief Executive Satya Nadella laid out the thinking behind the controls in an Oct. 10 essay"Source: [siliconangle.com](https://siliconangle.com/2026/10/19/microsoft-targets-the-agentic-insider-with-new-controls-for-rogue-ai-agents/), read 2026-11-16. - **microsoft-66** (third-party): On 2026-10-19, SiliconANGLE reported Microsoft gave no availability dates or pricing for the new agent controls announced that day.Quote: "Microsoft did not give availability dates or pricing for the new controls"Source: [siliconangle.com](https://siliconangle.com/2026/10/19/microsoft-targets-the-agentic-insider-with-new-controls-for-rogue-ai-agents/), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [Lasso Security profile](/vendors/lasso-security/) - [Noma Security profile](/vendors/noma-security/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # Noma Security AI runtime security URL: https://airuntimesecuritytools.com/vendors/noma-security/ Updated: 2026-11-16 Noma's AI-DR policies work on session context. Noma says a policy can block an action, mask sensitive data, steer the agent, bring in a human or trigger a killswitch, and that AI-DR detects prompt injection, tool poisoning, scope violations and drift from intent. Enforcement runs through gateways, agent hooks on developer machines, SDKs and APIs for cloud and SaaS agents, as SaaS or on-prem. Third-party press confirms Series A and Series B funding. Gaps from 16 Nov 2026: latency, fail-open or fail-closed behavior, named blocks of shell, file or package actions, public documentation and a price are none found, and credential theft is detection only. ## Strength and weakness for AI Runtime Security Tools **Strength:** Session-context runtime policy that can block, mask or steer agent actions (noma-security-20, noma-security-22). **Weakness:** No latency or fail-open/closed behavior and no named shell, file or package blocks; none found on 2026-11-16 (noma-security-38, noma-security-39, noma-security-50). ## Scores and evidence Noma Security is 16th in the overall ranking at 41.5 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Noma Security: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Action types stopped before they complete | 5 of 10 | 30 | The vendor claims blocking of tool calls and sensitive data and inspection for prompt injection, but credential theft is detection only. Blocking of shell, package, file and network actions: none found. | [noma-security-20](#fact-noma-security-20), [noma-security-23](#fact-noma-security-23), [noma-security-24](#fact-noma-security-24), [noma-security-22](#fact-noma-security-22), [noma-security-52](#fact-noma-security-52), [noma-security-50](#fact-noma-security-50) | | Layers covered | 6 of 10 | 20 | The vendor claims enforcement through gateways, agent hooks on developer machines, SDKs and APIs for cloud and SaaS agents. All of it is vendor-only, with limited endpoint action detail. | [noma-security-09](#fact-noma-security-09), [noma-security-18](#fact-noma-security-18), [noma-security-19](#fact-noma-security-19), [noma-security-41](#fact-noma-security-41) | | Latency and failure mode | 0 of 10 | 15 | Published inline latency and fail-open or fail-closed behavior: none found. | [noma-security-38](#fact-noma-security-38), [noma-security-39](#fact-noma-security-39) | | Deployment | 4 of 10 | 15 | The vendor says SaaS or on-prem, with enforcement through existing gateways and hooks. Supported operating systems and time to deploy: none found. | [noma-security-18](#fact-noma-security-18), [noma-security-33](#fact-noma-security-33), [noma-security-34](#fact-noma-security-34), [noma-security-29](#fact-noma-security-29) | | Detection and response | 4 of 10 | 5 | The vendor claims observability across prompts and tool calls, plus alerting. Named SIEM, SOAR or ticketing export: none found. | [noma-security-31](#fact-noma-security-31), [noma-security-21](#fact-noma-security-21), [noma-security-36](#fact-noma-security-36), [noma-security-35](#fact-noma-security-35) | | Transparency | 3 of 10 | 10 | The vendor publishes vulnerability research. Public documentation and a published price: none found. | [noma-security-44](#fact-noma-security-44), [noma-security-45](#fact-noma-security-45), [noma-security-42](#fact-noma-security-42), [noma-security-43](#fact-noma-security-43) | | Proof | 7 of 10 | 5 | Third-party press confirms Series A and Series B funding, and the vendor cites Gartner. Individually named customers: none found. | [noma-security-03](#fact-noma-security-03), [noma-security-54](#fact-noma-security-54), [noma-security-06](#fact-noma-security-06), [noma-security-05](#fact-noma-security-05) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 16th | 41.5 | | [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) | 16th | 44.5 | | [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) | 16th | 37.5 | | [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) | 16th | 48.0 | | [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) | 16th | 42.5 | | [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) | 16th | 42.5 | ## Product, deployment, price and owner - Product assessed: Noma AI-DR (AI Detection and Response) with Agent Access Control - Deployment: SaaS or on-prem; discovery via APIs, existing EDR or MDM; enforcement via hooks, gateways, SDKs and APIs - Time to deploy: none found on 2026-11-16 - Price: none found on 2026-11-16 - Owner: Noma Security (standalone company; no parent stated on its pages) - Website: [noma.security](https://noma.security) ## Facts and sources - **noma-security-01** (vendor): Noma Security describes its offering as an agent security and governance platform for enterprise security teams.Quote: "The Agent Security & Governance Platform that helps enterprise security teams enable agent adoption without losing control."Source: [noma.security](https://noma.security/), read 2026-11-16. - **noma-security-02** (vendor): Noma says it is an AI security platform built for enterprise AI adoption; its about page lists Niv Braun as Co-Founder and CEO and Alon Tron as Co-Founder and CTO.Quote: "Noma is the AI security platform purpose-built to secure enterprise AI adoption."Source: [noma.security](https://noma.security/about), read 2026-11-16. - **noma-security-03** (third-party): TechCrunch reported on 2024-10-31 that Noma, founded in 2023, closed a $25 million Series A led by Ballistic Ventures, for $32 million raised in total, after a $7 million seed.Quote: "Noma raised $7 million in a previously undisclosed seed round led by Glilot Capital Partners with participation from Cyber Club London."Source: [techcrunch.com](https://techcrunch.com/2024/10/31/noma-is-building-tools-to-spot-security-issues-with-ai-apps/), read 2026-11-16. - **noma-security-04** (vendor): In a press release dated 2026-10-28, Noma lists its backers as Evolution Equity Partners, Ballistic Ventures, Glilot Capital, Cyber Club London, Databricks Ventures and SVCI.Quote: "Backed by Evolution Equity Partners, Ballistic Ventures, Glilot Capital, Cyber Club London, Databricks Ventures, and SVCI"Source: [prnewswire.com](https://www.prnewswire.com/news-releases/noma-named-to-fourth-annual-fortune-cyber-60-list-of-venture-backed-cybersecurity-companies-302922012.html), read 2026-11-16. - **noma-security-05** (vendor): Noma says on its homepage that over 10% of the Fortune 100 trusts it; no individual customer names were shown on the pages read.Quote: "Trusted by over 10% of the Fortune 100"Source: [noma.security](https://noma.security/), read 2026-11-16. - **noma-security-06** (vendor): Noma says Gartner named it a Market Shaper in the 2026 Emerging Market Quadrant for AI Application Security, Startup Vendors.Quote: "Noma named a Market Shaper in the 2026 Gartner® Emerging Market Quadrant for AI Application Security, Startup Vendors"Source: [noma.security](https://noma.security/gartner/2026-gartner-market-quadrant), read 2026-11-16. - **noma-security-07** (vendor): On 2026-10-28 Noma announced it was named to the 2026-27 Fortune Cyber 60 list (early growth category), presented by Lightspeed Venture Partners and Amazon Web Services.Quote: "today announced it has been named to the 2026-27 Fortune Cyber 60, a list presented by Lightspeed Venture Partners and Amazon Web Services"Source: [prnewswire.com](https://www.prnewswire.com/news-releases/noma-named-to-fourth-annual-fortune-cyber-60-list-of-venture-backed-cybersecurity-companies-302922012.html), read 2026-11-16. - **noma-security-08** (vendor): Noma says it discovers endpoint agents, MCP servers and skills through the customer existing EDR or MDM, with no extra endpoint agent.Quote: "Noma discovers agents, MCP servers, and skills through your existing EDR or MDM. See where personal accounts and unapproved tools connect to company resources, without deploying another endpoint agent."Source: [noma.security](https://noma.security/platform/endpoint-agents), read 2026-11-16. - **noma-security-09** (vendor): On 2026-09-28 Noma announced endpoint agent security, saying enforcement coverage spans Claude Code, Claude Cowork, Cursor, Codex, Windsurf, Kiro, Antigravity and OpenClaw.Quote: "with coverage spanning agents including Claude Code, Claude Cowork, Cursor, Codex, Windsurf, Kiro, Antigravity, and OpenClaw"Source: [prnewswire.com](https://www.prnewswire.com/news-releases/noma-extends-agent-security-and-governance-to-the-employee-endpoint-302890925.html), read 2026-11-16. - **noma-security-18** (vendor): Noma says it enforces policy through AI gateways, MCP gateways, agent hooks, agent SDKs and direct APIs already in the customer environment.Quote: "Decouple governance from any single control point and enforce policy through AI gateways, MCP gateways, agent hooks, agent SDKs, and direct APIs, using the infrastructure already in your environment."Source: [noma.security](https://noma.security/solutions/mcp-security), read 2026-11-16. - **noma-security-19** (vendor): Noma says one policy set is enforced for agents on Bedrock behind an AI gateway, on Copilot Studio via integrations, and on developer machines via native hooks in Claude Code or Cursor.Quote: "The same behavioral monitoring and policy enforcement applies whether the agent runs on Bedrock behind an AI gateway"Source: [noma.security](https://noma.security/solutions/agent-behavior-runtime-control), read 2026-11-16. - **noma-security-20** (vendor): Noma says AI-DR policies can block an action, mask sensitive data, steer the agent, bring in a human or trigger a killswitch.Quote: "block an action, mask sensitive data, steer the agent towards a different path, get a human in the loop, or trigger a killswitch."Source: [noma.security](https://noma.security/products/ai-dr), read 2026-11-16. - **noma-security-21** (vendor): Noma says its response policy can alert, block or mask data, and that available enforcement actions depend on the integration.Quote: "Available enforcement actions depend on the integration."Source: [noma.security](https://noma.security/solutions/agent-behavior-runtime-control), read 2026-11-16. - **noma-security-22** (vendor): On 2026-09-28 Noma said AI-DR detects prompt injection, sensitive data leakage, malicious intent, tool poisoning, scope violations and drift from intent.Quote: "detects prompt injection, sensitive data leakage, malicious intent, tool poisoning, scope violations, and agents drifting from their intent or established behavior"Source: [prnewswire.com](https://www.prnewswire.com/news-releases/noma-extends-agent-security-and-governance-to-the-employee-endpoint-302890925.html), read 2026-11-16. - **noma-security-23** (vendor): Noma says it restricts individual tools in approved MCP servers and enforces the decision when agents attempt to connect or act.Quote: "Restrict individual tools within approved servers and enforce those decisions when agents attempt to connect or act."Source: [noma.security](https://noma.security/solutions/mcp-security), read 2026-11-16. - **noma-security-24** (vendor): Noma says it inspects MCP tool responses for hidden instructions, prompt injection, workflow hijacking and data exfiltration patterns; inspection of call arguments is not stated.Quote: "Noma inspects MCP tool responses for hidden instructions, prompt injection, workflow hijacking, and data exfiltration patterns."Source: [noma.security](https://noma.security/solutions/mcp-security), read 2026-11-16. - **noma-security-29** (none-found): None found on 2026-11-16: supported operating systems or a time to deploy, on the endpoint agents and MCP security pages.Source: [noma.security](https://noma.security/platform/endpoint-agents), read 2026-11-16. - **noma-security-31** (vendor): Noma says AI-DR gives observability into every prompt, tool call, response and interaction across endpoint, SaaS and cloud agents.Quote: "complete observability into every prompt, tool call, response, and interaction your agents engage in across endpoint, SaaS, and cloud environments"Source: [noma.security](https://noma.security/products/ai-dr), read 2026-11-16. - **noma-security-33** (vendor): Noma says it discovers SaaS agents through platform APIs with nothing to deploy.Quote: "Noma connects to SaaS agent platforms through their APIs and discovers every agent continuously, with nothing to deploy."Source: [noma.security](https://noma.security/platform/saas-agents), read 2026-11-16. - **noma-security-34** (vendor): Noma says it supports both on-prem and SaaS deployments.Quote: "Support for both on-prem and SaaS deployments, so no model, training data, or security events leave your environment."Source: [noma.security](https://noma.security/solutions/mcp-security), read 2026-11-16. - **noma-security-35** (none-found): None found on 2026-11-16: named SIEM, SOAR, ticketing or EDR export integrations (beyond EDR/MDM discovery) on the AI-DR page.Source: [noma.security](https://noma.security/products/ai-dr), read 2026-11-16. - **noma-security-36** (vendor): Noma says it has integrations with 80+ services; the services are not listed on the page read.Quote: "With integrations with 80+ services and coverage spanning endpoint, SaaS, and cloud agents"Source: [noma.security](https://noma.security/solutions/mcp-security), read 2026-11-16. - **noma-security-38** (none-found): None found on 2026-11-16: published latency of an inline decision on the AI-DR and runtime security pages.Source: [noma.security](https://noma.security/products/ai-dr), read 2026-11-16. - **noma-security-39** (none-found): None found on 2026-11-16: published fail-open or fail-closed behavior or timeouts on the runtime security page.Source: [noma.security](https://noma.security/solutions/agent-behavior-runtime-control), read 2026-11-16. - **noma-security-40** (vendor): Noma says AI Red Teaming runs multi-turn adversarial campaigns against agents and applications before release, with findings feeding runtime policies.Quote: "Noma AI Red Teaming runs multi-turn adversarial campaigns against your agents and applications before release"Source: [noma.security](https://noma.security/platform/homegrown-agents), read 2026-11-16. - **noma-security-41** (vendor): Noma says AI-DR evaluates each cloud agent action and, based on policy, alerts, blocks, masks data or routes to a human.Quote: "Based on policy, it alerts, blocks, masks data, or routes to a human."Source: [noma.security](https://noma.security/platform/homegrown-agents), read 2026-11-16. - **noma-security-42** (none-found): None found on 2026-11-16: a published price; noma.security/pricing returns Not Found and the sitemap lists no pricing page.Source: [noma.security](https://noma.security/pricing), read 2026-11-16. - **noma-security-43** (none-found): None found on 2026-11-16: public documentation; docs.noma.security redirects to a sign-in page.Source: [docs.noma.security](https://docs.noma.security), read 2026-11-16. - **noma-security-44** (vendor): Noma Labs published research on 2026-10-14 on Claude Code Desktop; the vendor says Anthropic scored the finding CVSS 8.7.Quote: "Anthropic scored Noma Labs' discovery as CVSS 8.7 (High) and treated the impact as equivalent to code execution on the host."Source: [noma.security](https://noma.security/noma-labs/sneakpeek-how-the-preview-pane-in-claude-code-desktop-can-exfiltrate-your-data), read 2026-11-16. - **noma-security-45** (vendor): Noma announced RufRoot, a CVSS 10.00 flaw its Labs found in the Ruflo agent platform (CVE-2026-59726); the release text read gives a 2026-06-30 publication date.Quote: "a maximum severity (CVSS 10.00) vulnerability discovered by Noma Labs in Ruflo"Source: [prnewswire.com](https://www.prnewswire.com/news-releases/noma-labs-discovers-critical-vulnerability-in-widely-adopted-open-source-ai-agent-platform-ruflo-302837494.html), read 2026-11-16. - **noma-security-49** (vendor): In its 2026-10-28 release Noma says it covers prebuilt agents like Claude Code, Cursor and GitHub Copilot running on developer machines.Quote: "prebuilt agents like Claude Code, Cursor, and GitHub Copilot running on developer machines"Source: [prnewswire.com](https://www.prnewswire.com/news-releases/noma-named-to-fourth-annual-fortune-cyber-60-list-of-venture-backed-cybersecurity-companies-302922012.html), read 2026-11-16. - **noma-security-50** (none-found): None found on 2026-11-16: named blocking of shell commands, package installs, file reads or writes, or network calls on the runtime security page.Source: [noma.security](https://noma.security/solutions/agent-behavior-runtime-control), read 2026-11-16. - **noma-security-51** (vendor): Noma says it flags the combination of untrusted input, sensitive data access and an external send as a session-level risk.Quote: "Untrusted input, sensitive data access, and an external send can each be fine alone and together form the lethal trifecta."Source: [noma.security](https://noma.security/solutions/agent-behavior-runtime-control), read 2026-11-16. - **noma-security-52** (vendor): Noma says it detects runtime behavior patterns that indicate credential theft, data exfiltration and AI memory tampering (detection, no block stated).Quote: "runtime detection of the behavioral patterns that indicate credential theft, data exfiltration, and AI memory tampering"Source: [prnewswire.com](https://www.prnewswire.com/news-releases/noma-labs-discovers-critical-vulnerability-in-widely-adopted-open-source-ai-agent-platform-ruflo-302837494.html), read 2026-11-16. - **noma-security-53** (vendor): On 2026-06-02 Noma said AI-DR detects prompt injection, data exfiltration and scope violations in real time.Quote: "It detects prompt injection, data exfiltration, and scope violations in real time."Source: [prnewswire.com](https://www.prnewswire.com/news-releases/noma-launches-agentic-access-control-to-govern-ai-agents-and-mcp-servers-across-the-enterprise-302788534.html), read 2026-11-16. - **noma-security-54** (third-party): SiliconANGLE reported on 2025-07-31 that Noma raised a $100 million Series B led by Evolution Equity Partners with Ballistic Ventures and Glilot Capital.Quote: "Evolution Equity Partners led the Series B investment with participation from Ballistic Ventures and Glilot Capital."Source: [siliconangle.com](https://siliconangle.com/2025/07/31/ai-security-startup-noma-security-closes-100m-funding-round/), read 2026-11-16. ## News mentioning Noma Security - [Noma extends agent runtime protection to the endpoint](/news/noma-endpoint-agent-security/) ## Related pages - [All vendor profiles](/vendors/) - [Microsoft profile](/vendors/microsoft/) - [Operant AI profile](/vendors/operant-ai/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # Operant AI AI runtime security URL: https://airuntimesecuritytools.com/vendors/operant-ai/ Updated: 2026-11-16 Operant names its blocks in action terms. It says Tool Intent Guard stops exfiltration, bulk dumps, credential access and unauthorized sharing before execution, and that endpoint clients block shell calls, package installs and pasted secrets. Enforcement runs at the endpoint, browser, LLM gateway and agent layers, rolled out by MDM in under 15 minutes or by a Helm install in under five, with VPC and air-gapped options. A 0.7 ms P99 figure appears only in a demo panel, so it counts as a mention. Gaps from 16 Nov 2026: a latency SLA, fail-open or fail-closed behavior, public documentation, a price and named paying customers are none found. ## Strength and weakness for AI Runtime Security Tools **Strength:** Names blocked action types before execution (shell, packages, credentials, bulk reads, payments) across agent and endpoint layers (operant-ai-24, operant-ai-26, operant-ai-27, operant-ai-31). **Weakness:** No published inline latency SLA or fail-open/fail-closed behavior: none found on 2026-11-16; only a 0.7ms demo figure (operant-ai-32, operant-ai-33). ## Scores and evidence Operant AI is joint 4th in the overall ranking at 62.5 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Operant AI: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Action types stopped before they complete | 7 of 10 | 30 | The vendor claims blocking of shell calls, package installs, pasted secrets, credential access, untrusted tool use and prompt-injected commands before execution. All of it is vendor-only. | [operant-ai-26](#fact-operant-ai-26), [operant-ai-27](#fact-operant-ai-27), [operant-ai-28](#fact-operant-ai-28), [operant-ai-24](#fact-operant-ai-24), [operant-ai-21](#fact-operant-ai-21), [operant-ai-30](#fact-operant-ai-30), [operant-ai-29](#fact-operant-ai-29) | | Layers covered | 7 of 10 | 20 | The vendor claims endpoint, browser, LLM gateway and agent enforcement. It notes the Copilot path detects but is not inline, and all of it is vendor-only. | [operant-ai-09](#fact-operant-ai-09), [operant-ai-13](#fact-operant-ai-13), [operant-ai-22](#fact-operant-ai-22), [operant-ai-23](#fact-operant-ai-23), [operant-ai-15](#fact-operant-ai-15) | | Latency and failure mode | 4 of 10 | 15 | A 0.7 ms P99 figure appears only in a demo panel. Fail-open or fail-closed behavior: none found. | [operant-ai-32](#fact-operant-ai-32), [operant-ai-33](#fact-operant-ai-33) | | Deployment | 7 of 10 | 15 | Native clients rolled out by MDM in under 15 minutes, a helm install in under 5 minutes, and VPC, on-prem and air-gapped options. All of it is vendor-claimed. | [operant-ai-10](#fact-operant-ai-10), [operant-ai-34](#fact-operant-ai-34), [operant-ai-35](#fact-operant-ai-35) | | Detection and response | 7 of 10 | 5 | The vendor claims telemetry forwarding to named SIEMs, audit logs with decision reasons, and activity timelines. All of it is vendor-only. | [operant-ai-36](#fact-operant-ai-36), [operant-ai-38](#fact-operant-ai-38), [operant-ai-37](#fact-operant-ai-37), [operant-ai-40](#fact-operant-ai-40) | | Transparency | 4 of 10 | 10 | Research is confirmed by third-party press and observed open-source code. Public documentation and a published price: none found. | [operant-ai-41](#fact-operant-ai-41), [operant-ai-42](#fact-operant-ai-42), [operant-ai-43](#fact-operant-ai-43), [operant-ai-44](#fact-operant-ai-44) | | Proof | 7 of 10 | 5 | A third-party startup list and an observed marketplace listing, plus vendor-stated funding and testimonials. Named paying customers and confirmed analyst coverage: none found. | [operant-ai-05](#fact-operant-ai-05), [operant-ai-07](#fact-operant-ai-07), [operant-ai-03](#fact-operant-ai-03), [operant-ai-08](#fact-operant-ai-08) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | joint 4th | 62.5 | | [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) | 5th | 65.5 | | [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) | 6th | 61.0 | | [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) | joint 4th | 67.0 | | [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) | 5th | 65.5 | | [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) | joint 5th | 65.5 | ## Product, deployment, price and owner - Product assessed: Semantic Firewall, Agent Protector (with CodeInjectionGuard and ScopeGuard), Endpoint Protector and AI Gatekeeper - Deployment: Endpoint clients rolled out by MDM; Kubernetes helm install for cloud and API protection; gateway for LLM and MCP traffic; VPC, on-prem and air-gapped options on Enterprise - Time to deploy: Vendor says endpoint rollout in under 15 minutes and helm install in under 5 minutes - Price: No prices published; three plans (Pro, Scale, Enterprise) with a 7-day trial; none found on 2026-11-16 for amounts - Owner: Operant AI, Inc. (standalone company, San Francisco) - Website: [operant.ai](https://www.operant.ai) ## Facts and sources - **operant-ai-01** (vendor): Operant AI describes itself as a runtime defense platform for live cloud and AI applications; its footer names Operant AI, Inc. and San Francisco.Quote: "Operant AI is the only Runtime AI Application Defense Platform that actively protects every layer of live cloud and AI applications"Source: [operant.ai](https://www.operant.ai/company/about), read 2026-11-16. - **operant-ai-02** (vendor): The homepage footer shows the owner as Operant AI, Inc., a standalone company, with SOC 2 Type II compliance stated.Quote: "Operant is SOC 2, Type II Compliant."Source: [operant.ai](https://www.operant.ai), read 2026-11-16. - **operant-ai-03** (vendor): Operant says its Series A of $10 million, led by SineWave Ventures and Felicis, brought total funding to $13.5 million (post dated 2024).Quote: "Operant's $10 million Series A brings Operant's total funding to $13.5 million."Source: [operant.ai](https://www.operant.ai/art-kubed/operant-ai-series-a), read 2026-11-16. - **operant-ai-04** (vendor): Operant's about page lists Vrajesh Bhavsar as Co-Founder and CEO and Priyanka Tembey as Co-Founder and CTO.Quote: "Co-Founder and CEO"Source: [operant.ai](https://www.operant.ai/company/about), read 2026-11-16. - **operant-ai-05** (third-party): CRN's 2026-07-02 list of the 10 hottest AI security startups includes Operant AI, founded 2021, CEO Vrajesh Bhavsar.Quote: "Operant AI Founded: 2021 CEO: Vrajesh Bhavsar"Source: [crn.com](https://www.crn.com/news/security/2026/the-10-hottest-ai-security-startups-of-2026-so-far?page=5), read 2026-11-16. - **operant-ai-06** (vendor): Operant says it is the only vendor featured across five Gartner AI security reports (vendor claim, not on its own checked).Quote: "Operant AI is the only vendor featured across all five of Gartner®'s most critical AI security reports"Source: [operant.ai](https://www.operant.ai), read 2026-11-16. - **operant-ai-07** (observed): Operant's AWS Marketplace seller profile exists and repeats its runtime defense positioning; the page lists Operant as an AWS Marketplace seller.Quote: "Operant AI is the world's only Runtime Application Defense Platform."Source: [aws.amazon.com](https://aws.amazon.com/marketplace/seller-profile?id=seller-gebxlkhuqojto), read 2026-11-16. - **operant-ai-08** (vendor): The homepage shows attributed testimonials (for example Chargebee, Juniper Networks, Cohere) but no list of named paying customers; none found on 2026-11-16 beyond testimonial attributions.Quote: "Cyber Security Leader | Chargebee"Source: [operant.ai](https://www.operant.ai), read 2026-11-16. - **operant-ai-09** (vendor): Operant says Endpoint Protector inspects AI IDEs, coding agents, desktop clients, MCP clients, skills, tools and plugins on employee endpoints (generally available).Quote: "Endpoint Protector inspects every AI IDE, coding agent, desktop client, MCP client, skill, tool, and plugin"Source: [operant.ai](https://www.operant.ai/platform/endpoint-protector), read 2026-11-16. - **operant-ai-10** (vendor): Operant says Endpoint Protector ships native clients for macOS, Windows and Linux, rolled out by MDM (Jamf, Intune, Workspace ONE, Kandji), with workforce rollout in under 15 minutes.Quote: "Workforce-wide rollout in under 15 minutes."Source: [operant.ai](https://www.operant.ai/platform/endpoint-protector), read 2026-11-16. - **operant-ai-12** (vendor): Operant has a browser protection page for AI browsers and gen AI websites, enforced by a local engine on the device.Quote: "Your browser has access to everything you're logged into. Now it takes orders from the pages it reads."Source: [operant.ai](https://www.operant.ai/solutions/browser-protection), read 2026-11-16. - **operant-ai-13** (vendor): On 2026-08-27, Operant announced Browser Protection that puts an allow, sanitize or block decision inline on the browser, in both directions, with the same policy engine and audit trail.Quote: "Operant Browser Protection puts the same allow/sanitize/block decision inline on that surface, in both directions"Source: [operant.ai](https://www.operant.ai/art-kubed/introducing-operant-semantic-firewall), read 2026-11-16. - **operant-ai-15** (vendor): On 2026-10-01, Operant announced inline defense for custom agents built on the Microsoft 365 Agents SDK plus Purview-based observability for Copilot; it states the Purview path detects but does not sit inline.Quote: "Agent Protector goes inside the agent loop of custom Copilot agents, so each step is evaluated before it executes."Source: [operant.ai](https://www.operant.ai/art-kubed/announcing-operants-coverage-for-microsoft-copilot-and-every-ai-assistant-embedded-in-microsoft-365), read 2026-11-16. - **operant-ai-21** (vendor): Operant says MCP Gateway builds MCP trust zones, blocks untrusted servers and tools in real time, and enforces least privilege on tool use.Quote: "Implement MCP trust zones with real-time blocking of untrusted servers and tools."Source: [operant.ai](https://www.operant.ai/solutions/mcp-gateway), read 2026-11-16. - **operant-ai-22** (vendor): Operant says AI Gatekeeper is an LLM gateway with a native firewall, multi-provider routing and request inspection for AI applications.The exact quote is on the source page.Source: [operant.ai](https://www.operant.ai/platform/pricing), read 2026-11-16. - **operant-ai-23** (vendor): Operant says Agent Protector blocks inline when an agent goes outside established patterns, escalates privileges or shows data exfiltration.Quote: "Agent Protector blocks the action inline."Source: [operant.ai](https://www.operant.ai/platform/agent-protector), read 2026-11-16. - **operant-ai-24** (vendor): Operant says Semantic Firewall's Tool Intent Guard blocks exfiltration, bulk dumps, credential access and unauthorized sharing before execution.Quote: "blocks exfiltration, bulk dumps, credential access, and unauthorized sharing"Source: [operant.ai](https://www.operant.ai/platform/semantic-firewall), read 2026-11-16. - **operant-ai-25** (vendor): On 2026-08-27, Operant launched Semantic Firewall, which returns allow, block or redact decisions as the agent acts, with a plain-language reason for each; no numeric latency is given on that page.Quote: "returns an allow, block, or redact decision as the agent acts"Source: [operant.ai](https://www.operant.ai/art-kubed/introducing-operant-semantic-firewall), read 2026-11-16. - **operant-ai-26** (vendor): Operant says its Code Intent Guard inspects generated commands, scripts and shell calls before execution and blocks shell breakout and privilege escalation.Quote: "Inspects generated commands, scripts, and shell calls pre-execution"Source: [operant.ai](https://www.operant.ai/platform/semantic-firewall), read 2026-11-16. - **operant-ai-27** (vendor): Operant says Endpoint Protector's CodeInjectionGuard blocks malicious payloads, prompt-injected commands and unsanctioned package installs before they run.Quote: "blocking malicious payloads, prompt-injected commands, and unsanctioned package installs before they run"Source: [operant.ai](https://www.operant.ai/platform/endpoint-protector), read 2026-11-16. - **operant-ai-28** (vendor): Operant says Endpoint Protector blocks pasted source code, API keys and credentials before they reach the model, with inline redaction of 100+ data types.Quote: "Blocks pasted IP, API keys, and credentials before they hit the model"Source: [operant.ai](https://www.operant.ai/platform/endpoint-protector), read 2026-11-16. - **operant-ai-29** (vendor): Operant says CodeInjectionGuard (launched in a post dated 2026-04-21) scans packages that agents pull at runtime and blocks malicious code before it executes.Quote: "detects and blocks malicious code before it is executed by endpoint AI agents"Source: [operant.ai](https://www.operant.ai/art-kubed/introducing-operant-codeinjectionguard-for-ai-agents), read 2026-11-16. - **operant-ai-30** (vendor): Operant says Agent ScopeGuard blocks out-of-scope actions before execution, limiting the data sources and APIs an agent may use.Quote: "blocking out-of-scope actions before execution and alerting your security team with full action traces"Source: [operant.ai](https://www.operant.ai/platform/agent-protector), read 2026-11-16. - **operant-ai-31** (vendor): On 2026-10-15, Operant said Agent Protector's Agentic Commerce Defense evaluates each agent payment attempt and can block a payment before it is submitted.Quote: "Operant can block the payment before it is submitted."Source: [operant.ai](https://www.operant.ai/art-kubed/stop-dollar-exfiltration-operant-agent-protector-now-defends-agentic-commerce), read 2026-11-16. - **operant-ai-32** (vendor): The Endpoint Protector page shows a '0.7ms P99 Inspection' figure inside a live-inspection demo panel; it is a page display, not a stated SLA.Quote: "0.7ms P99 Inspection"Source: [operant.ai](https://www.operant.ai/platform/endpoint-protector), read 2026-11-16. - **operant-ai-33** (none-found): None found on 2026-11-16: published fail-open or fail-closed behavior or timeout settings on the Semantic Firewall page, Endpoint Protector page or pricing page.Source: [operant.ai](https://www.operant.ai/platform/semantic-firewall), read 2026-11-16. - **operant-ai-34** (vendor): Operant says its platform deploys with a single-step helm install on Kubernetes 1.25+, works in under 5 minutes, and needs zero instrumentation.Quote: "Single step helm install. Zero instrumentation. Zero Integrations. Works in <5 minutes."Source: [operant.ai](https://www.operant.ai/platform/faqs), read 2026-11-16. - **operant-ai-35** (vendor): Operant's Enterprise plan lists VPC, on-prem and air-gapped options plus SSO, SCIM, MDM templates and API/webhooks.Quote: "VPC, on-prem, and air-gapped options"Source: [operant.ai](https://www.operant.ai/platform/pricing), read 2026-11-16. - **operant-ai-36** (vendor): Operant says Endpoint Protector forwards agent-loop telemetry to Splunk, Datadog, Elastic, Chronicle and more.Quote: "Forwards rich telemetry to Splunk, Datadog, Elastic, Chronicle, and more"Source: [operant.ai](https://www.operant.ai/platform/endpoint-protector), read 2026-11-16. - **operant-ai-37** (vendor): Operant says Semantic Firewall policies are written in natural language and every decision carries a plain-language reason for audit.Quote: "Every decision carries a plain-language reason that stands up to review."Source: [operant.ai](https://www.operant.ai/platform/semantic-firewall), read 2026-11-16. - **operant-ai-38** (vendor): Operant's pricing lists standard audit logs in Pro and SIEM, SOC/SOAR integrations and extended retention from the Scale plan.Quote: "SIEM, SOC/SOAR integrations"Source: [operant.ai](https://www.operant.ai/platform/pricing), read 2026-11-16. - **operant-ai-40** (vendor): Operant says Agent Protector traces agents from prompts through tool calls to memory access, with activity timelines and tool activity graphs.Quote: "comprehensive tracing from initial prompts through tool calls to memory store access"Source: [operant.ai](https://www.operant.ai/platform/agent-protector), read 2026-11-16. - **operant-ai-41** (vendor): The product pages are readable without sign-in, but the pricing page lists plans without prices and says to talk to the vendor.Quote: "All plans support multi-year commitments with volume discounts. Talk to us for details."Source: [operant.ai](https://www.operant.ai/platform/pricing), read 2026-11-16. - **operant-ai-42** (none-found): None found on 2026-11-16: public product documentation. docs.operant.ai redirects to the product sign-in page.Source: [docs.operant.ai](https://docs.operant.ai), read 2026-11-16. - **operant-ai-43** (observed): Operant publishes Woodpecker, an open-source red teaming tool, on GitHub under the Apache-2.0 license.Quote: "woodpecker is a modular red teaming tool focused for AI and cloud apps."Source: [github.com](https://github.com/OperantAI/woodpecker), read 2026-11-16. - **operant-ai-44** (third-party): GBHackers (October 2025) reports Operant's research team found the Shadow Escape zero-click MCP attack and that Operant reported it to OpenAI and filed a CVE.Quote: "Operant AI has reported this attack to OpenAI and filed a CVE to address this emerging threat to data governance and privacy."Source: [gbhackers.com](https://gbhackers.com/zero-click-exploit-targets-ai-agents/), read 2026-11-16. - **operant-ai-45** (third-party): CyberTech Intelligence reported on the Endpoint Protector launch (Operant's news page dates it 2026-05-04) as a way to secure AI interactions at the endpoint.Quote: "Operant AI has launched Endpoint Protector, a new solution designed to secure AI interactions directly at the endpoint."Source: [cybertechnologyinsights.com](https://cybertechnologyinsights.com/endpoint-security/operant-ai-introduces-endpoint-security-for-ai-tools/), read 2026-11-16. ## News mentioning Operant AI - [Operant launches Semantic Firewall for agent actions](/news/operant-semantic-firewall-launch/) ## Related pages - [All vendor profiles](/vendors/) - [Noma Security profile](/vendors/noma-security/) - [Palo Alto Networks profile](/vendors/palo-alto-networks/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # Palo Alto Networks AI runtime security URL: https://airuntimesecuritytools.com/vendors/palo-alto-networks/ Updated: 2026-11-16 Palo Alto Networks sells AI runtime security as Prisma AIRS, and its strength is inline scanning of content. The API Intercept scan checks prompts and model responses for any model, and an AI security profile sets an allow, alert or block action that stops the request and records a violation. Network Intercept and gateway modes add coverage without an endpoint agent, and partner docs show prompt gating in coding agents. Documents are public. Gaps from 16 Nov 2026: tool calls are scanned only when the agent calls the scanner, and blocking of shell, file, package, network and credential actions on the machine, a millisecond latency figure and fail-open or fail-closed behavior are none found. ## Strength and weakness for AI Runtime Security Tools **Strength:** Inline prompt, response and tool-call scanning with allow, alert and block actions at the API, network and gateway layers (palo-alto-networks-47, palo-alto-networks-53, palo-alto-networks-57). **Weakness:** None found on 2026-11-16 for published inline latency, fail-open or fail-closed behavior, or blocking of shell, file, package or credential actions (palo-alto-networks-64, palo-alto-networks-69). ## Scores and evidence Palo Alto Networks is 7th in the overall ranking at 58.5 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Palo Alto Networks: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Action types stopped before they complete | 6 of 10 | 30 | Blocking of sensitive uploads, prompt injection and data leakage is documented, and prompt gating is observed in partner docs. Tool calls are scanned only when the agent calls the scanner, and on-machine action blocking: none found. | [palo-alto-networks-20](#fact-palo-alto-networks-20), [palo-alto-networks-42](#fact-palo-alto-networks-42), [palo-alto-networks-58](#fact-palo-alto-networks-58), [palo-alto-networks-47](#fact-palo-alto-networks-47), [palo-alto-networks-59](#fact-palo-alto-networks-59), [palo-alto-networks-45](#fact-palo-alto-networks-45), [palo-alto-networks-69](#fact-palo-alto-networks-69) | | Layers covered | 7 of 10 | 20 | The model, application and agent layers are documented and partly observed. Endpoint enforcement is a separate profile, and the observed hooks inspect prompts only. | [palo-alto-networks-41](#fact-palo-alto-networks-41), [palo-alto-networks-53](#fact-palo-alto-networks-53), [palo-alto-networks-45](#fact-palo-alto-networks-45), [palo-alto-networks-51](#fact-palo-alto-networks-51), [palo-alto-networks-59](#fact-palo-alto-networks-59), [palo-alto-networks-60](#fact-palo-alto-networks-60) | | Latency and failure mode | 4 of 10 | 15 | An observed verdict timeout setting is documented. A millisecond latency figure and fail-open or fail-closed behavior: none found. | [palo-alto-networks-61](#fact-palo-alto-networks-61), [palo-alto-networks-62](#fact-palo-alto-networks-62), [palo-alto-networks-64](#fact-palo-alto-networks-64) | | Deployment | 6 of 10 | 15 | Network and API intercept modes, a SaaS or hybrid gateway and API limits are documented. Published time to deploy for runtime security: none found. | [palo-alto-networks-52](#fact-palo-alto-networks-52), [palo-alto-networks-44](#fact-palo-alto-networks-44), [palo-alto-networks-54](#fact-palo-alto-networks-54), [palo-alto-networks-55](#fact-palo-alto-networks-55) | | Detection and response | 4 of 10 | 5 | Logs of verdicts, ticket creation and false positive reporting are documented. SIEM or SOAR export: none found. | [palo-alto-networks-46](#fact-palo-alto-networks-46), [palo-alto-networks-66](#fact-palo-alto-networks-66), [palo-alto-networks-65](#fact-palo-alto-networks-65), [palo-alto-networks-67](#fact-palo-alto-networks-67) | | Transparency | 6 of 10 | 10 | Docs pages and a research report are public, and example hooks are observed on GitHub. Published price: none found. | [palo-alto-networks-26](#fact-palo-alto-networks-26), [palo-alto-networks-50](#fact-palo-alto-networks-50), [palo-alto-networks-54](#fact-palo-alto-networks-54), [palo-alto-networks-24](#fact-palo-alto-networks-24) | | Proof | 7 of 10 | 5 | A listed company with named customers and a vendor-cited Gartner recognition. The only third-party fact covers a partner integration, so standalone confirmation of customers, funding or analyst coverage: none found. | [palo-alto-networks-28](#fact-palo-alto-networks-28), [palo-alto-networks-29](#fact-palo-alto-networks-29), [palo-alto-networks-30](#fact-palo-alto-networks-30), [palo-alto-networks-27](#fact-palo-alto-networks-27), [palo-alto-networks-32](#fact-palo-alto-networks-32) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 7th | 58.5 | | [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) | 8th | 58.5 | | [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) | 7th | 56.5 | | [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) | joint 8th | 60.5 | | [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) | 10th | 54.0 | | [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) | 7th | 59.0 | ## Product, deployment, price and owner - Product assessed: Prisma AIRS AI Runtime Security (API Intercept, Network Intercept, AI Gateway) - Deployment: AI Access Security via existing NGFW and SSE gateways with no endpoint agent; Prisma Browser as a browser, extension or mobile app; Prisma AIRS via API Intercept, Network Intercept and AI Gateway (SaaS or Hybrid) - Time to deploy: none found on 2026-11-16 (Prisma Browser: access 'in minutes' per vendor) - Price: none found on 2026-11-16 - Owner: Palo Alto Networks, Inc. (NASDAQ: PANW) - Website: paloaltonetworks.com ## Facts and sources - **palo-alto-networks-01** (vendor): The Palo Alto Networks leadership page lists Nikesh Arora as chairman and CEO, so the company is its own parent and the owner of Prisma AIRS, AI Access Security and Prisma Browser.Quote: "Nikesh Arora is chairman and CEO of Palo Alto Networks, the world’s leading cybersecurity company."Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/about-us/management), read 2026-11-16. - **palo-alto-networks-02** (vendor): On 2026-05-29 Palo Alto Networks said it closed the acquisition of Portkey, an AI gateway company, and made it the AI Gateway of Prisma AIRS.Quote: "today announced it has closed the acquisition of Portkey, a pioneer in AI Gateways."Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/company/press/2026/palo-alto-networks-completes-acquisition-of-portkey-to-secure-ai-agents), read 2026-11-16. - **palo-alto-networks-03** (vendor): Palo Alto Networks documents Prisma AIRS as its AI security platform covering AI applications, models, agents and infrastructure across the AI lifecycle.Quote: "Prisma® AIRS™ is Palo Alto Networks' comprehensive AI security platform."Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/ai-runtime-security), read 2026-11-16. - **palo-alto-networks-04** (none-found): None found on 2026-11-16: endpoint-agent coverage for desktop or command-line agents inside the Prisma AIRS product page; the page lists Agentic Endpoint Security as a separate capability, which is a separate profile and not counted here.Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/prisma/prisma-ai-runtime-security), read 2026-11-16. - **palo-alto-networks-14** (vendor): Palo Alto Networks says AI Access Security deploys through existing SSE and firewall gateways with no additional endpoint agents or infrastructure changes.The exact quote is on the source page.Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/sase/ai-access-security), read 2026-11-16. - **palo-alto-networks-16** (vendor): Palo Alto Networks documents Prisma Browser as a browser, an extension for consumer browsers and a mobile app, with access in minutes.Quote: "Available as a browser, an extension for consumer browsers, and a mobile app, it provides secure access from any location in minutes."Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/prisma-browser), read 2026-11-16. - **palo-alto-networks-17** (observed): The Prisma Browser deployment docs describe Windows MSI installers for x64 and ARM64 and a macOS installation, with weekly updates.Quote: "Prisma Browser supports two installation variations for macOS environments."Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-browser/deployment), read 2026-11-16. - **palo-alto-networks-20** (vendor): Palo Alto Networks says AI Access Security inline detection blocks sensitive text and file transfers to GenAI apps.Quote: "Inline data detection ensures regulatory compliance and blocks sensitive text- and file-based data transfer to GenAI apps."Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/sase/ai-access-security), read 2026-11-16. - **palo-alto-networks-24** (none-found): None found on 2026-11-16: a published price for Prisma AIRS, AI Access Security or Prisma Browser on their product pages; guessed pricing URLs on paloaltonetworks.com returned 404.Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/sase/ai-access-security), read 2026-11-16. - **palo-alto-networks-26** (vendor): Palo Alto Networks lists a published research report on prompt attacks among the AI Access Security resources.Quote: "Securing GenAI: A Comprehensive Report on Prompt Attacks: Taxonomy, Risks and Solutions"Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/sase/ai-access-security), read 2026-11-16. - **palo-alto-networks-27** (vendor): Palo Alto Networks says Gartner recognized it as a Market Shaper in the Emerging Market Quadrant for AI Application Security in September 2026.The exact quote is on the source page.Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/prisma/prisma-ai-runtime-security), read 2026-11-16. - **palo-alto-networks-28** (vendor): Palo Alto Networks says it is a NASDAQ-listed company with more than 70,000 customers.Quote: "Trusted by 70,000+ customers and powered by Unit 42 threat intelligence"Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/company/press/2026/palo-alto-networks-completes-acquisition-of-portkey-to-secure-ai-agents), read 2026-11-16. - **palo-alto-networks-29** (vendor): Prisma AIRS lists Glean's CISO as a named customer, quoted on the product page.Quote: "What makes Palo Alto Networks AI Runtime stand out are three things."Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/prisma/prisma-ai-runtime-security), read 2026-11-16. - **palo-alto-networks-30** (vendor): Prisma AIRS lists Moveworks and TELUS Digital as named customers, quoted on the product page.Quote: "A product like Prisma AIRS is not just good for Moveworks. It's absolutely essential for the industry."Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/prisma/prisma-ai-runtime-security), read 2026-11-16. - **palo-alto-networks-32** (third-party): On 2026-08-20 SiliconANGLE reported that Zero Networks expanded its Palo Alto Networks integration, with an AI Segmentation link to Prisma AIRS that has no release date.Quote: "The Prisma AIRS integration and Windows traffic redirection has not been given a release date."Source: [siliconangle.com](https://siliconangle.com/2026/08/20/zero-networks-expands-palo-alto-networks-integration-to-ai-agent-control/), read 2026-11-16. - **palo-alto-networks-41** (vendor): The Prisma AIRS docs say AI Gateway sits between users or applications and every LLM provider they call.Quote: "sits between your users and applications and every large language model (LLM) provider they call."Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/prisma-airs/ai-gateway/ai-gateway-overview), read 2026-11-16. - **palo-alto-networks-42** (vendor): The Prisma AIRS docs say AI Gateway guardrails run inline on every request and response, blocking prompt injection and detecting sensitive data.Quote: "Guardrails run inline on every request and response, blocking prompt injection attempts, detecting sensitive data, and enforcing topic restrictions"Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/prisma-airs/ai-gateway/ai-gateway-overview), read 2026-11-16. - **palo-alto-networks-44** (vendor): The AI Gateway docs offer a SaaS model and a Hybrid model where the data plane runs in the customer's own Kubernetes cluster.Quote: "Hybrid , where you host the data plane in your own Kubernetes cluster so that request and response payloads never leave your environment."Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/prisma-airs/ai-gateway/ai-gateway-overview), read 2026-11-16. - **palo-alto-networks-45** (vendor): The Prisma AIRS MCP Server docs say it intercepts tool invocations, scans them and returns a verdict, as a tool the agent is instructed to call.Quote: "The Prisma AIRS MCP Server operates by intercepting tool invocations, performing security analysis, and then returning a verdict on whether a threat was detected."Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-agent-identity/understanding-the-prisma-airs-mcp-server.html), read 2026-11-16. - **palo-alto-networks-46** (vendor): The Prisma AIRS MCP Server docs say detailed logs track tool invocations, security verdicts and detected threats.Quote: "You can monitor these interactions through detailed logs that track tool invocations, security verdicts, and any detected threats."Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-agent-identity/understanding-the-prisma-airs-mcp-server.html), read 2026-11-16. - **palo-alto-networks-47** (vendor): The AI Security Profile docs say a Block action intercepts and stops the request and records a violation.Quote: "Block: Intercepts and stops the request, records a violation at the configured severity"Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/configure-allow-alert-block-actions-ai-security-profiles.html), read 2026-11-16. - **palo-alto-networks-48** (vendor): Palo Alto Networks says Prisma AIRS agent security addresses memory manipulation and tool misuse.Quote: "It provides dedicated AI Agent Security against risks like memory manipulation and tool misuse, ensuring safe interactions with tools and workflows."Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/prisma/prisma-ai-runtime-security), read 2026-11-16. - **palo-alto-networks-50** (observed): A Palo Alto Networks GitHub repo holds community example hooks for coding assistants (Claude Code, Codex, Cursor, Cline, Gemini CLI and others), supported on a best-effort basis.Quote: "supported as best effort by Palo Alto Networks"Source: [github.com](https://github.com/PaloAltoNetworks/prisma-airs-integrations), read 2026-11-16. - **palo-alto-networks-51** (observed): The same GitHub repo says Prisma AIRS scans prompts, responses and tool interactions, and lists gateway and coding assistant examples.Quote: "Prisma AIRS provides inline security for AI applications, scanning prompts, responses, and tool interactions in real-time."Source: [github.com](https://github.com/PaloAltoNetworks/prisma-airs-integrations), read 2026-11-16. - **palo-alto-networks-52** (vendor): Palo Alto Networks says Prisma AIRS offers Network Intercept and API Intercept deployment modes.Quote: "It offers deployment modes such as Network Intercept and API Intercept to secure AI applications, AI models, AI data, and AI agents."Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/prisma/prisma-ai-runtime-security), read 2026-11-16. - **palo-alto-networks-53** (vendor): The API Intercept docs say a Scan API checks prompts and model responses and works with any model.Quote: "The Scan API service scans prompts and models responses to identify potential threats and provides actionable recommendations."Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/ai-runtime-security-api-intercept-overview.html), read 2026-11-16. - **palo-alto-networks-54** (vendor): The API Intercept docs give limits of 2 MB per synchronous scan request and one API key per deployment profile.Quote: "2 MB maximum payload size per synchronous scan request"Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/ai-runtime-security-api-intercept-overview.html), read 2026-11-16. - **palo-alto-networks-55** (vendor): The AI Runtime Security docs describe an instance deployed in the customer's cloud environment that inspects traffic between AI apps, models and datasets.Quote: "An Prisma AIRS instance is a specific deployment of AI Runtime Security within your cloud environment."Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/prisma-airs/getting-started/ai-runtime-security-overview), read 2026-11-16. - **palo-alto-networks-56** (vendor): The AI Runtime Security docs name prompt injection, sensitive data leakage, insecure output and model DoS as threats it defends against.Quote: "from network attack threats such as prompt injections, sensitive data leakage, insecure output (for example, malware and URLs), and model DoS attacks."Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/prisma-airs/getting-started/ai-runtime-security-overview), read 2026-11-16. - **palo-alto-networks-57** (vendor): Palo Alto Networks says Prisma AIRS monitors AI traffic and detects prompt injection, sensitive data exposure, malicious URLs and toxic content.Quote: "detecting threats like prompt injection, sensitive data exposure, malicious URLs, and toxic content before they impact your AI workflows."Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/prisma/prisma-ai-runtime-security), read 2026-11-16. - **palo-alto-networks-58** (vendor): The API Intercept docs list blocking sensitive data leaks during AI interactions as a use case.Quote: "Prevent sensitive data leakage : Use API-based threat detection to block sensitive data leaks during AI interactions."Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/ai-runtime-security-api-intercept-overview.html), read 2026-11-16. - **palo-alto-networks-59** (observed): The Anthropic inference hooks docs say Anthropic pauses inference, sends the prompt to Prisma AIRS and enforces its allow or deny verdict.Quote: "Anthropic pauses inference, POSTs the conversation transcript to a Prisma AIRS endpoint, and enforces the allow or deny verdict AIRS returns"Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/integrate-anthropic-inference-hooks.html), read 2026-11-16. - **palo-alto-networks-60** (observed): The same docs list limits: prompt events only, US region only, and tool_use blocks carried but not inspected.Quote: "Prompt event only. Response-side enforcement is not yet supported."Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/integrate-anthropic-inference-hooks.html), read 2026-11-16. - **palo-alto-networks-61** (observed): The inference hooks docs set a verdict timeout of 1 to 10,000 ms and warn that heavy profiles add latency.The exact quote is on the source page.Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/integrate-anthropic-inference-hooks.html), read 2026-11-16. - **palo-alto-networks-62** (vendor): The hooks docs recommend lightweight profiles for latency-sensitive gating and say DLP and malicious-code detection add latency.Quote: "Lightweight profiles (prompt injection, topic guardrails) are recommended for latency-sensitive gating; heavy profiles (DLP, malicious-code detection) add latency."Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/integrate-anthropic-inference-hooks.html), read 2026-11-16. - **palo-alto-networks-63** (vendor): The Prisma AIRS page lists an October 2026 update on inspecting streamed LLM responses to remove latency trade-offs.Quote: "Protect AI apps that stream LLM responses using, eliminating latency trade-offs."Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/prisma/prisma-ai-runtime-security), read 2026-11-16. - **palo-alto-networks-64** (none-found): None found on 2026-11-16: a published inline decision latency in milliseconds, or fail-open or fail-closed behavior, on the API Intercept overview page and the Anthropic hooks page.Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/ai-runtime-security-api-intercept-overview.html), read 2026-11-16. - **palo-alto-networks-65** (vendor): The Prisma AIRS page lists a November 2026 update adding false positive and false negative reporting and auto-remediation for API Intercept.Quote: "Introducing FP/FN Reporting and Auto-Remediation for AI Runtime Security API Intercept."Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/prisma/prisma-ai-runtime-security), read 2026-11-16. - **palo-alto-networks-66** (vendor): The SaaS Agent Security docs say security teams can open actionable tickets directly from findings.Quote: "enabling security teams to create actionable tickets directly from security findings."Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/prisma-airs/ai-agent-identity/saas-agent-security-overview), read 2026-11-16. - **palo-alto-networks-67** (none-found): None found on 2026-11-16: SIEM or SOAR export, or IdP and EDR integrations, in the Prisma AIRS docs pages read (overview, API Intercept, gateway, SaaS agent security).Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/ai-runtime-security), read 2026-11-16. - **palo-alto-networks-68** (vendor): The AI Security Profile docs list Allow, Alert and Block actions per detector, with Alert recording a violation without blocking.Quote: "Alert: Permits traffic to pass while recording a violation and generating a log at the configured severity."Source: [docs.paloaltonetworks.com](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/configure-allow-alert-block-actions-ai-security-profiles.html), read 2026-11-16. - **palo-alto-networks-69** (none-found): None found on 2026-11-16: blocking of shell commands, file access, package installs or credential access by Prisma AIRS on the machine in the pages read; the vendor puts endpoint enforcement in the separate Agentic Endpoint Security profile.Source: [paloaltonetworks.com](https://www.paloaltonetworks.com/prisma/prisma-ai-runtime-security), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [Operant AI profile](/vendors/operant-ai/) - [Pillar Security profile](/vendors/pillar-security/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # Pillar Security AI runtime security URL: https://airuntimesecuritytools.com/vendors/pillar-security/ Updated: 2026-11-16 Pillar covers all four layers: model, application, agent and endpoint. It says its endpoint product blocks reads of keys, .env files, SSH and cloud credentials from reaching external destinations, blocks persistence git hooks and stops agents from rewriting their own configuration, with endpoint decisions run locally and offline. Runtime guardrails block prompt injection, jailbreaking and tool manipulation, and an activity view shows blocked decisions on session timelines with SIEM streaming. Installation is an MDM script on macOS, Linux and Windows. Gaps from 16 Nov 2026: a latency figure, fail-open behavior, public documentation and a price are none found. Vanta announced a definitive agreement to acquire Pillar on 9 Nov 2026. ## Strength and weakness for AI Runtime Security Tools **Strength:** Covers model, application, agent and endpoint layers, blocking credential exfiltration and prompt injection (pillar-security-10, pillar-security-25). **Weakness:** None found on 2026-11-16 for published decision latency or fail-open behavior, and public docs sit behind sign-in (pillar-security-32, pillar-security-35). ## Scores and evidence Pillar Security is 8th in the overall ranking at 55.5 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Pillar Security: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Action types stopped before they complete | 7 of 10 | 30 | The vendor claims pre-execution decisions in coding-agent hooks, with blocking of credential and file reads, persistence writes, prompt injection, tool manipulation and sensitive data. All of it is vendor-only. | [pillar-security-09](#fact-pillar-security-09), [pillar-security-10](#fact-pillar-security-10), [pillar-security-25](#fact-pillar-security-25), [pillar-security-26](#fact-pillar-security-26) | | Layers covered | 7 of 10 | 20 | The vendor claims endpoint hooks, agent pre-tool decisions, model guardrails and AI gateway integration. All of it is vendor-only. | [pillar-security-09](#fact-pillar-security-09), [pillar-security-04](#fact-pillar-security-04), [pillar-security-25](#fact-pillar-security-25), [pillar-security-27](#fact-pillar-security-27) | | Latency and failure mode | 0 of 10 | 15 | The vendor says endpoint decisions run locally and offline, but a latency figure and fail-open or fail-closed behavior: none found. | [pillar-security-32](#fact-pillar-security-32), [pillar-security-11](#fact-pillar-security-11) | | Deployment | 6 of 10 | 15 | MDM script install on macOS, Linux and Windows, with managed cloud, self-hosted or hybrid options and go-live steps. Time to deploy: none found. | [pillar-security-06](#fact-pillar-security-06), [pillar-security-29](#fact-pillar-security-29), [pillar-security-30](#fact-pillar-security-30), [pillar-security-33](#fact-pillar-security-33) | | Detection and response | 7 of 10 | 5 | The vendor claims an activity view of blocked decisions with session timelines, an audit trail and SIEM streaming. All of it is vendor-only. | [pillar-security-12](#fact-pillar-security-12), [pillar-security-24](#fact-pillar-security-24), [pillar-security-28](#fact-pillar-security-28) | | Transparency | 4 of 10 | 10 | Third-party press confirms the vendor's published research. Public documentation and a published price: none found. | [pillar-security-36](#fact-pillar-security-36), [pillar-security-42](#fact-pillar-security-42), [pillar-security-43](#fact-pillar-security-43), [pillar-security-34](#fact-pillar-security-34), [pillar-security-35](#fact-pillar-security-35) | | Proof | 8 of 10 | 5 | Third-party press covers the seed funding and a Fortune 300 proof of concept, and the vendor shows named testimonials and a Gartner recognition. Large named customers: none found. | [pillar-security-39](#fact-pillar-security-39), [pillar-security-41](#fact-pillar-security-41), [pillar-security-37](#fact-pillar-security-37), [pillar-security-40](#fact-pillar-security-40) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 8th | 55.5 | | [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) | 7th | 60.5 | | [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) | 9th | 50.0 | | [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) | 7th | 64.0 | | [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) | 6th | 61.0 | | [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) | 8th | 58.5 | ## Product, deployment, price and owner - Product assessed: Runtime Guardrails and Agentic Endpoint (Runtime Guardian) - Deployment: Managed cloud, self-hosted or hybrid; endpoint product by MDM script, with a per-user service and agent hooks - Time to deploy: none found on 2026-11-16 - Price: none found on 2026-11-16 - Owner: Pillar Security. Vanta announced a definitive agreement to acquire it on 9 Nov 2026; the deal was not stated as closed on 16 Nov 2026 - Website: [pillar.security](https://www.pillar.security) ## Facts and sources - **pillar-security-01** (vendor): On 2026-11-09 Pillar's founders announced that Vanta has signed a definitive agreement to acquire Pillar Security, so the deal is announced and not described as closed.Quote: "Today, Vanta has signed a definitive agreement to acquire Pillar Security."Source: [pillar.security](https://www.pillar.security/blog/vanta-to-acquire-pillar-security-building-trust-in-ai-together), read 2026-11-16. - **pillar-security-02** (vendor): Pillar says it was founded in 2023 and describes its product as discovering, testing and securing AI applications and agents and blocking unauthorized actions at runtime.Quote: "When we founded Pillar in 2023, most of the conversation around AI security was about chatbots and prompts."Source: [pillar.security](https://www.pillar.security/blog/vanta-to-acquire-pillar-security-building-trust-in-ai-together), read 2026-11-16. - **pillar-security-03** (vendor): Pillar says its platform discovers, tests, secures and governs AI agents, across five modules: AI Discovery & Posture, Red Teaming & Attack Surface Exposure, Runtime Guardrails, Agentic Endpoint, and Governance & Compliance.Quote: "One platform to discover, test, secure and govern every AI agent across your organization"Source: [pillar.security](https://www.pillar.security), read 2026-11-16. - **pillar-security-04** (vendor): Pillar says its Agentic Endpoint product finds coding agents, MCP servers, skills and plugins on every managed machine.Quote: "Find coding agents, MCP servers, skills and plugins on every machine."Source: [pillar.security](https://www.pillar.security/platform/agentic-endpoint), read 2026-11-16. - **pillar-security-06** (vendor): Pillar says the endpoint product is deployed through MDM scripts (Jamf Pro, Kandji, JumpCloud or any MDM that runs scripts) on macOS, Linux and Windows.Quote: "Roll out with Jamf Pro, Kandji, JumpCloud or any MDM that runs scripts, on macOS, Linux and Windows."Source: [pillar.security](https://www.pillar.security/platform/agentic-endpoint), read 2026-11-16. - **pillar-security-09** (vendor): Pillar says Runtime Guardian sits in the pre-tool hook of Claude Code, Codex CLI and Cursor and decides on every action before it runs.Quote: "Runtime Guardian sits in the agent's own pre-tool hook for Claude Code, Codex CLI and Cursor, and decides on every action before it runs."Source: [pillar.security](https://www.pillar.security/platform/agentic-endpoint), read 2026-11-16. - **pillar-security-10** (vendor): Pillar says it blocks reads of keys, .env, SSH and cloud credentials from reaching external destinations, and blocks persistence git hooks and agents rewriting their own configuration.Quote: "Stop reads of keys, .env, SSH and cloud credentials from reaching external destinations."Source: [pillar.security](https://www.pillar.security/platform/agentic-endpoint), read 2026-11-16. - **pillar-security-11** (vendor): Pillar says endpoint decisions run locally and keep working offline, and that on macOS and Linux MDM installs managed hooks for Codex and Cursor that developers cannot disable.Quote: "Decisions run locally and keep working offline; on macOS and Linux, MDM installs managed hooks for Codex and Cursor that developers cannot disable."Source: [pillar.security](https://www.pillar.security/platform/agentic-endpoint), read 2026-11-16. - **pillar-security-12** (vendor): Pillar says each control can be set to Off, Monitor or Block, and every blocked or asked decision appears in an Activity view with the full session timeline.Quote: "Every decision, blocked or asked, appears in Activity with the full session timeline."Source: [pillar.security](https://www.pillar.security/platform/agentic-endpoint), read 2026-11-16. - **pillar-security-24** (vendor): Pillar says its audit trail logs every model deployment, policy decision and compliance check with metadata.Quote: "Log every model deployment, policy decision, and compliance check with complete metadata."Source: [pillar.security](https://www.pillar.security/platform/governance-compliance), read 2026-11-16. - **pillar-security-25** (vendor): Pillar says its runtime guardrails detect and block prompt injection, jailbreaking and tool manipulation attempts.Quote: "Detect and block prompt injection, jailbreaking, and tool manipulation attempts."Source: [pillar.security](https://www.pillar.security/platform/runtime-guardrails), read 2026-11-16. - **pillar-security-26** (vendor): Pillar says runtime guardrails mask, block or log PII, PHI, secrets and credentials in AI interactions according to configurable policies.Quote: "Identify, mask, or block PII, PHI, secrets, and credentials in AI interactions."Source: [pillar.security](https://www.pillar.security/platform/runtime-guardrails), read 2026-11-16. - **pillar-security-27** (vendor): Pillar lists AI gateways it works with (LiteLLM, Kong, TrueFoundry, Agent Router, Open WebUI) and says any gateway works via API.Quote: "+ any gateway via API"Source: [pillar.security](https://www.pillar.security/platform/runtime-guardrails), read 2026-11-16. - **pillar-security-28** (vendor): Pillar says it logs prompts, responses, tool calls and results with timestamps, and streams telemetry to SIEM platforms.Quote: "Stream telemetry to security platforms for centralized monitoring."Source: [pillar.security](https://www.pillar.security/platform/runtime-guardrails), read 2026-11-16. - **pillar-security-29** (vendor): Pillar says it runs as managed cloud, self-hosted in the customer's cloud account or data center, or hybrid.Quote: "Deploy Pillar where your data lives: our managed cloud, your own cloud account or data center, or a mix of both."Source: [pillar.security](https://www.pillar.security/platform/deployment), read 2026-11-16. - **pillar-security-30** (vendor): Pillar's go-live steps are scope, install, connect (AI gateway, code repositories, device management) and go live in monitor mode before blocking.Quote: "Start in monitor mode, then switch to blocking."Source: [pillar.security](https://www.pillar.security/platform/deployment), read 2026-11-16. - **pillar-security-31** (vendor): Pillar says adaptive guardrails are available in managed cloud only today, and that self-hosted runs core detection models in the customer environment.Quote: "Adaptive guardrails are available in our managed cloud only today. (as of October 2026)"Source: [pillar.security](https://www.pillar.security/platform/deployment), read 2026-11-16. - **pillar-security-32** (none-found): None found on 2026-11-16: published latency of an inline decision or fail-open or fail-closed behavior, searched on the runtime guardrails, agentic endpoint and deployment pages.Source: [pillar.security](https://www.pillar.security/platform/runtime-guardrails), read 2026-11-16. - **pillar-security-33** (none-found): None found on 2026-11-16: a published time to deploy or time to first inventory, searched on the deployment, agentic endpoint and discovery pages.Source: [pillar.security](https://www.pillar.security/platform/deployment), read 2026-11-16. - **pillar-security-34** (none-found): None found on 2026-11-16: a published price; the pricing URL returned 404 and the site offers only a demo request.Source: [pillar.security](https://www.pillar.security/pricing), read 2026-11-16. - **pillar-security-35** (none-found): None found on 2026-11-16: public product documentation; docs.pillar.security loads a sign-in page marked noindex.Source: [docs.pillar.security](https://docs.pillar.security), read 2026-11-16. - **pillar-security-36** (vendor): Pillar publishes the SAIL framework and a research blog and says SAIL has passed 50,000 downloads and maps 91 AI risks across seven lifecycle phases.Quote: "SAIL has passed 50,000 downloads, and maps 91 AI risks across seven lifecycle phases."Source: [pillar.security](https://www.pillar.security/blog/vanta-to-acquire-pillar-security-building-trust-in-ai-together), read 2026-11-16. - **pillar-security-37** (vendor): Pillar's site shows testimonials from named people at Eleos, another company and Similarweb.The exact quote is on the source page.Source: [pillar.security](https://www.pillar.security/platform/agentic-endpoint), read 2026-11-16. - **pillar-security-38** (vendor): Pillar says a Fortune 300 manufacturer runs it fully self-hosted.Quote: "A Fortune 300 manufacturer runs Pillar fully self-hosted."Source: [pillar.security](https://www.pillar.security/platform/deployment), read 2026-11-16. - **pillar-security-39** (third-party): CRN reported that Pillar announced a $9 million seed round led by Shield Capital in April 2025, and listed it among its 10 hottest AI security startups of 2026.Quote: "Pillar announced its $9 million seed funding round led by Shield Capital in April 2025."Source: [crn.com](https://www.crn.com/news/security/2026/the-10-hottest-ai-security-startups-of-2026-so-far?page=6), read 2026-11-16. - **pillar-security-40** (vendor): The Pillar newsroom lists a Gartner Cool Vendor in AI Software Security recognition dated Jul 1, 2026 and a Frost & Sullivan award dated Sep 17, 2025.Quote: "Pillar Security Named as a 2026 Gartner® Cool Vendor in AI Software Security"Source: [pillar.security](https://www.pillar.security/newsroom), read 2026-11-16. - **pillar-security-41** (third-party): A Refresh Miami profile dated 2026-01-26 says Pillar scanned more than 30,000 repositories in a Fortune 300 proof of concept and found 8,000 AI models and 400 AI systems unknown to the security team.Quote: "Pillar scanned more than 30,000 code repositories. The result: 8,000 AI models and 400 AI systems"Source: [refreshmiami.com](https://refreshmiami.com/news/ai-broke-the-security-model-pillar-wants-to-fix-it/), read 2026-11-16. - **pillar-security-42** (third-party): BleepingComputer reported on 2026-07-20 that Pillar researchers published a Week of Sandbox Escapes series on sandbox bypasses in AI coding agents, with findings sorted into four failure modes.Quote: "Pillar sorts the seven findings into four failure modes:"Source: [bleepingcomputer.com](https://www.bleepingcomputer.com/news/security/cursor-codex-gemini-cli-antigravity-hit-by-sandbox-escapes/), read 2026-11-16. - **pillar-security-43** (third-party): SecurityWeek reported on 2026-05-07 that Pillar disclosed a critical Gemini CLI vulnerability that could have led to supply chain compromise via a GitHub issue.Quote: "A critical vulnerability in Gemini CLI could have allowed attackers to mount a supply chain attack via indirect prompts injected into a GitHub issue, Pillar Security warns."Source: [securityweek.com](https://www.securityweek.com/gemini-cli-vulnerability-could-have-led-to-code-execution-supply-chain-attack/), read 2026-11-16. - **pillar-security-44** (vendor): The Pillar newsroom lists a press release dated Oct 13, 2026 announcing BadEngram research on supply-chain risk in open-weight AI models.Quote: "Pillar Security Discovers ‘BadEngram,’ Exposing Hidden Supply-Chain Risk in Open-Weight AI Models"Source: [pillar.security](https://www.pillar.security/newsroom), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [Palo Alto Networks profile](/vendors/palo-alto-networks/) - [SentinelOne (Prompt Security) profile](/vendors/sentinelone/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # SentinelOne (Prompt Security) AI runtime security URL: https://airuntimesecuritytools.com/vendors/sentinelone/ Updated: 2026-11-16 SentinelOne offers AI runtime control through Prompt Security, which it agreed to acquire in August 2025. SentinelOne says it blocks adversarial prompts and scrubs sensitive outputs with a real-time AI firewall, and third-party press confirms prompt blocking and inline redaction. A self-hosted on-premises version, announced on 23 Mar 2026, works in disconnected environments. Delivery is a browser extension, a dedicated AI endpoint agent, an MCP gateway or API integrations. Gaps from 16 Nov 2026: published inline latency, fail-open or fail-closed behavior, and blocking of shell, file, network, credential or package actions are none found, and the agent gateway integration was announced as upcoming. ## Strength and weakness for AI Runtime Security Tools **Strength:** Real-time AI firewall for prompts and outputs, plus agent-layer enforcement and a self-hosted option (sentinelone-41, sentinelone-43, sentinelone-46). **Weakness:** None found on 2026-11-16 for published inline latency, fail-open or fail-closed behavior, or blocking of shell, file, network or credential actions (sentinelone-51, sentinelone-53). ## Scores and evidence SentinelOne (Prompt Security) is 14th in the overall ranking at 45.5 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **SentinelOne (Prompt Security): score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Action types stopped before they complete | 5 of 10 | 30 | Prompt blocking and inline redaction are confirmed by third-party press, and the vendor claims prompt injection and output scrubbing. Shell, file, network, credential and package install blocking: none found. | [sentinelone-19](#fact-sentinelone-19), [sentinelone-21](#fact-sentinelone-21), [sentinelone-41](#fact-sentinelone-41), [sentinelone-43](#fact-sentinelone-43), [sentinelone-52](#fact-sentinelone-52), [sentinelone-53](#fact-sentinelone-53) | | Layers covered | 6 of 10 | 20 | The vendor claims endpoint agent, AI app firewall and agent gateway coverage, with third-party mention of MCP gateway security. The agent gateway integration was announced as upcoming. | [sentinelone-32](#fact-sentinelone-32), [sentinelone-41](#fact-sentinelone-41), [sentinelone-35](#fact-sentinelone-35), [sentinelone-40](#fact-sentinelone-40), [sentinelone-34](#fact-sentinelone-34) | | Latency and failure mode | 0 of 10 | 15 | Published latency figure and fail-open or fail-closed behavior: none found. | [sentinelone-51](#fact-sentinelone-51), [sentinelone-54](#fact-sentinelone-54) | | Deployment | 6 of 10 | 15 | The vendor says it deploys in minutes through a browser extension, endpoint agent and MCP gateway, or on-prem. OS list and deployment requirements: none found. | [sentinelone-17](#fact-sentinelone-17), [sentinelone-13](#fact-sentinelone-13), [sentinelone-14](#fact-sentinelone-14), [sentinelone-46](#fact-sentinelone-46), [sentinelone-18](#fact-sentinelone-18) | | Detection and response | 4 of 10 | 5 | The vendor claims audit trails and searchable session logs. Named SIEM export, SOAR or ticketing: none found, and its own SIEM link is upcoming. | [sentinelone-22](#fact-sentinelone-22), [sentinelone-44](#fact-sentinelone-44), [sentinelone-36](#fact-sentinelone-36), [sentinelone-48](#fact-sentinelone-48) | | Transparency | 4 of 10 | 10 | Research and an observed open-source repo are published. Price for this product and public docs: none found. | [sentinelone-27](#fact-sentinelone-27), [sentinelone-28](#fact-sentinelone-28), [sentinelone-25](#fact-sentinelone-25), [sentinelone-26](#fact-sentinelone-26) | | Proof | 7 of 10 | 5 | Third-party press covers funding and the acquisition, with named CISO quotes. Analyst coverage is only vendor-linked. | [sentinelone-03](#fact-sentinelone-03), [sentinelone-04](#fact-sentinelone-04), [sentinelone-05](#fact-sentinelone-05), [sentinelone-29](#fact-sentinelone-29), [sentinelone-30](#fact-sentinelone-30), [sentinelone-31](#fact-sentinelone-31) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 14th | 45.5 | | [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) | 14th | 48.0 | | [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) | 13th | 41.0 | | [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) | 13th | 51.5 | | [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) | 15th | 45.0 | | [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) | joint 12th | 50.0 | ## Product, deployment, price and owner - Product assessed: Prompt Security AI Application Security and AI firewall, plus agent and endpoint enforcement (SentinelOne) - Deployment: Singularity Platform console; browser extension, dedicated AI endpoint agent, MCP Gateway and API integrations; self-hosted Prompt Security On-Premises; AgentCore gateway integration - Time to deploy: SentinelOne says it deploys in minutes; no measured figure or OS list found on 2026-11-16 - Price: none found on 2026-11-16 for Prompt Security; the pricing page lists Singularity tiers and says Enterprise is Contact Sales for Pricing - Owner: SentinelOne, Inc. (NYSE: S), Mountain View, California - Website: sentinelone.com ## Facts and sources - **sentinelone-01** (vendor): SentinelOne says Prompt Security is built into its Singularity Platform and managed from the same console as its endpoint, cloud, identity and data security.Quote: "Prompt Security is built into the SentinelOne Singularity Platform."Source: [sentinelone.com](https://www.sentinelone.com/platform/securing-ai-prompt/), read 2026-11-16. - **sentinelone-02** (vendor): On 2025-08-05 SentinelOne announced a definitive agreement to acquire Prompt Security, so the AI security line sits inside SentinelOne (NYSE: S).Quote: "signed a definitive agreement to acquire Prompt Security"Source: [sentinelone.com](https://www.sentinelone.com/press/sentinelone-to-acquire-prompt-security-to-advance-genai-security/), read 2026-11-16. - **sentinelone-03** (third-party): CRN reports that Prompt Security was founded in 2023 and raised an $18 million Series A in November 2024 before SentinelOne agreed to acquire it.Quote: "Founded in 2023, Prompt Security offers protection against AI-related threats"Source: [crn.com](https://www.crn.com/news/security/2025/sentinelone-to-acquire-ai-protection-startup-prompt-security/), read 2026-11-16. - **sentinelone-04** (third-party): CRN reports the Series A was led by Jump Capital with Hetz Ventures, Ridge Ventures, Okta and F5 taking part.Quote: "In November 2024, Prompt Security announced an $18 million Series A funding round led by Jump Capital"Source: [crn.com](https://www.crn.com/news/security/2025/sentinelone-to-acquire-ai-protection-startup-prompt-security/), read 2026-11-16. - **sentinelone-05** (third-party): Techzine reports SentinelOne acquired Israeli company Prompt Security to prepare its Singularity platform for secure AI use in organizations.Quote: "SentinelOne has announced the acquisition of Israeli company Prompt Security."Source: [techzine.eu](https://www.techzine.eu/news/security/133547/sentinelone-strengthens-ai-protection-with-prompt-security-deal/), read 2026-11-16. - **sentinelone-06** (third-party): Wikipedia lists SentinelOne as a public company headquartered in Mountain View, California, founded in 2013, with Tomer Weingarten as chairman and CEO.Quote: "Headquarters Mountain View, California , U.S."Source: [en.wikipedia.org](https://en.wikipedia.org/wiki/SentinelOne), read 2026-11-16. - **sentinelone-07** (vendor): SentinelOne says Prompt Security inventories every AI tool and code assistant in use, including unsanctioned shadow AI.Quote: "Inventory every AI tool and code assistant in use, including unsanctioned shadow AI"Source: [sentinelone.com](https://www.sentinelone.com/platform/securing-ai-prompt/), read 2026-11-16. - **sentinelone-13** (vendor): Prompt Security says AI governance needs local sensors, namely a browser extension and a dedicated AI endpoint agent, because network controls alone fall short.Quote: "you need local, purpose-built sensors: a browser extension and a dedicated AI endpoint agent"Source: [prompt.security](https://prompt.security/blog/the-key-layer-in-ai-security-browser-and-endpoint-sensors), read 2026-11-16. - **sentinelone-14** (vendor): SentinelOne says an MCP Gateway, browser extension, security plugins and API integrations work together so no agent surface goes undetected.Quote: "MCP Gateway, browser extension, security plugins, and API integrations work together so no agent surface goes undetected"Source: [sentinelone.com](https://www.sentinelone.com/platform/prompt-security-agentic-ai-security/), read 2026-11-16. - **sentinelone-15** (vendor): In its acquisition release SentinelOne says Prompt Security's technology integrates across browsers, desktop applications and APIs.Quote: "integrating across browsers, desktop applications, and APIs"Source: [sentinelone.com](https://www.sentinelone.com/press/sentinelone-to-acquire-prompt-security-to-advance-genai-security/), read 2026-11-16. - **sentinelone-17** (vendor): SentinelOne says Prompt Security deploys in minutes with no complex integrations or infrastructure changes.Quote: "Prompt Security deploys in minutes. No complex integrations or infrastructure changes are required."Source: [sentinelone.com](https://www.sentinelone.com/platform/securing-ai-prompt/), read 2026-11-16. - **sentinelone-18** (none-found): None found on 2026-11-16: looked on the Prompt Security product pages for supported operating systems, browsers, a measured time to first AI inventory and a published deployment requirements list; only "minutes" was found.Source: [sentinelone.com](https://www.sentinelone.com/platform/securing-ai-prompt/), read 2026-11-16. - **sentinelone-19** (vendor): SentinelOne says AI Usage Control redacts PII, secrets and other sensitive data inline on the device before a prompt leaves it.Quote: "Redact PII, secrets, and other sensitive data at the edge, inline, before a prompt ever leaves the device"Source: [sentinelone.com](https://www.sentinelone.com/platform/prompt-security-ai-usage-control/), read 2026-11-16. - **sentinelone-21** (third-party): Techzine reports that Prompt Security uses policy-based controls, for example blocking risky prompts before they are sent.Quote: "Policy-based controls enforce safe AI usage by, for example, blocking risky prompts before they are sent."Source: [techzine.eu](https://www.techzine.eu/news/security/133547/sentinelone-strengthens-ai-protection-with-prompt-security-deal/), read 2026-11-16. - **sentinelone-22** (vendor): SentinelOne says AI Usage Control keeps a full audit trail of every prompt, redaction and policy action.Quote: "Review every prompt, redaction, and policy action in one place, with a full audit trail across every AI interaction"Source: [sentinelone.com](https://www.sentinelone.com/platform/prompt-security-ai-usage-control/), read 2026-11-16. - **sentinelone-25** (vendor): SentinelOne's public pricing page lists tiers but says the Enterprise tier is "Contact Sales for Pricing" and does not list Prompt Security.Quote: "Contact Sales for Pricing"Source: [sentinelone.com](https://www.sentinelone.com/pricing/), read 2026-11-16. - **sentinelone-26** (none-found): None found on 2026-11-16: looked on the pricing page and Prompt Security pages for a published price of Prompt Security or AI Usage Control and for public technical documentation readable without sign-in; found marketing pages and a demo form only.Source: [sentinelone.com](https://www.sentinelone.com/pricing/), read 2026-11-16. - **sentinelone-27** (vendor): On 2026-08-20 Prompt Security published research on three agent cases (AGENTS.md goal hijack, marketplace skills dependency hijack, instruction file integrity) and says approval prompts should show the real destination and payload.Quote: "approval prompts that show the real destination and payload instead of a friendly summary"Source: [prompt.security](https://prompt.security/blog/when-ai-agents-become-the-supply-chain-hidden-control-planes-in-agentic-systems), read 2026-11-16. - **sentinelone-28** (observed): Prompt Security publishes ClawSec on GitHub under AGPL-3.0, a security skill suite for OpenClaw, Hermes, PicoClaw and NanoClaw agents with 1.1k stars.Quote: "ClawSec source code is licensed under GNU AGPL-3.0-or-later"Source: [github.com](https://github.com/prompt-security/clawsec), read 2026-11-16. - **sentinelone-29** (vendor): SentinelOne's page quotes Margarita Rivera, Global CISO of Carnival Corporation, on Prompt Security giving a first layer of visibility and protection.Quote: "Prompt Security has been incredible for our organization."Source: [sentinelone.com](https://www.sentinelone.com/platform/prompt-security-ai-usage-control/), read 2026-11-16. - **sentinelone-30** (vendor): SentinelOne's page quotes Richard Moore, CISO at 10x Banking, on Prompt Security.Quote: "Prompt Security has allowed us to be enablers for the business."Source: [sentinelone.com](https://www.sentinelone.com/platform/prompt-security-ai-usage-control/), read 2026-11-16. - **sentinelone-31** (vendor): SentinelOne's Prompt Security page links a "2026 Latio AI Security Market Report" among its resources (analyst-style coverage cited by the vendor; contents not read).Quote: "2026 Latio AI Security Market Report"Source: [sentinelone.com](https://www.sentinelone.com/platform/securing-ai-prompt/), read 2026-11-16. - **sentinelone-32** (vendor): Prompt Security says a dedicated AI endpoint agent should detect new agents, inspect prompt construction, enforce runtime policies and prevent sensitive data exposure before execution.Quote: "Prevent sensitive data exposure before execution"Source: [prompt.security](https://prompt.security/blog/the-key-layer-in-ai-security-browser-and-endpoint-sensors), read 2026-11-16. - **sentinelone-34** (vendor): On 2026-06-17 SentinelOne announced an upcoming Prompt Security integration with Amazon Bedrock AgentCore supplying signals for prompt injection, PII exposure, tool-use validation, LLM response monitoring and data leakage prevention.Quote: "prompt injection detection, PII exposure, tool-use validation, LLM response monitoring, and data leakage prevention"Source: [sentinelone.com](https://www.sentinelone.com/press/sentinelone-announces-ai-security-integration-with-amazon-bedrock-agentcore-to-deliver-runtime-guardrails-for-ai-agents/), read 2026-11-16. - **sentinelone-35** (vendor): SentinelOne says AgentCore acts on Prompt Security signals at the gateway, outside the agent's reasoning loop, with deterministic allow-or-deny enforcement.Quote: "Because detection can be probabilistic, but enforcement stays deterministic"Source: [sentinelone.com](https://www.sentinelone.com/press/sentinelone-announces-ai-security-integration-with-amazon-bedrock-agentcore-to-deliver-runtime-guardrails-for-ai-agents/), read 2026-11-16. - **sentinelone-36** (vendor): On 2026-08-04 SentinelOne said it would connect Prompt Security, Singularity Cloud Security and Singularity AI SIEM in a single view across Amazon Bedrock and AgentCore, with general availability targeted for AWS re:Invent 2026.Quote: "Prompt Security, Singularity™ Cloud Security, and Singularity™ AI SIEM into a single view across Amazon Bedrock"Source: [sentinelone.com](https://www.sentinelone.com/press/sentinelone-expands-collaboration-with-aws-to-deliver-unified-ai-governance/), read 2026-11-16. - **sentinelone-40** (third-party): Techzine reports Prompt Security works standalone of the AI model and via the Model Context Protocol through MCP gateway security.Quote: "also works via the now established Model Context Protocol via MCP gateway security"Source: [techzine.eu](https://www.techzine.eu/news/security/133547/sentinelone-strengthens-ai-protection-with-prompt-security-deal/), read 2026-11-16. - **sentinelone-41** (vendor): SentinelOne says Prompt Security blocks adversarial prompts and scrubs sensitive outputs with a real-time AI firewall for AI apps.Quote: "Block adversarial prompts and scrub sensitive outputs with a real-time AI firewall"Source: [sentinelone.com](https://www.sentinelone.com/platform/securing-ai-prompt/), read 2026-11-16. - **sentinelone-42** (vendor): SentinelOne says Prompt Security enforces policy at the point of interaction, not after the fact.Quote: "Policies enforced at the point of interaction, not after the fact."Source: [sentinelone.com](https://www.sentinelone.com/platform/securing-ai-prompt/), read 2026-11-16. - **sentinelone-43** (vendor): SentinelOne says Prompt Security stops agent policy violations before they propagate across connected systems.Quote: "Stop violations before they propagate across connected systems"Source: [sentinelone.com](https://www.sentinelone.com/platform/prompt-security-agentic-ai-security/), read 2026-11-16. - **sentinelone-44** (vendor): SentinelOne says Prompt Security produces a searchable log of every session, action and tool an agent touches.Quote: "Produce a searchable log of every session, action, and tool an agent touches."Source: [sentinelone.com](https://www.sentinelone.com/platform/prompt-security-agentic-ai-security/), read 2026-11-16. - **sentinelone-46** (vendor): On 2026-03-23 SentinelOne said Prompt Security On-Premises is self-hosted, works in disconnected environments, and blocks prompt injections and data leaks as a firewall for internal and external AI tools.Quote: "Prompt Security blocks threats like prompt injections and data leaks"Source: [sentinelone.com](https://www.sentinelone.com/press/sentinelone-brings-ai-security-to-on-premise-regulated-sovereign-self-hosted-and-airgapped-environments/), read 2026-11-16. - **sentinelone-47** (vendor): SentinelOne says customers with existing Prompt Security licenses can use them with AgentCore through a Bring Your Own License path.Quote: "Customers with existing Prompt Security licenses will apply them through a Bring Your Own License path."Source: [sentinelone.com](https://www.sentinelone.com/press/sentinelone-announces-ai-security-integration-with-amazon-bedrock-agentcore-to-deliver-runtime-guardrails-for-ai-agents/), read 2026-11-16. - **sentinelone-48** (none-found): None found on 2026-11-16: looked on the Prompt Security pages and press releases for a named SIEM export (beyond Singularity AI SIEM), SOAR, EDR, identity provider or ticketing integration for Prompt Security events; none named.Source: [sentinelone.com](https://www.sentinelone.com/platform/securing-ai-prompt/), read 2026-11-16. - **sentinelone-50** (vendor): On 2026-02-06 SentinelOne described its AI Security platform as covering AI-SPM, runtime workload protection, employee GenAI security and agent security.Quote: "employee GenAI security and agent security"Source: [sentinelone.com](https://www.sentinelone.com/press/sentinelone-delivers-end-to-end-ai-security-from-data-to-runtime/), read 2026-11-16. - **sentinelone-51** (none-found): None found on 2026-11-16: looked on the Prompt Security, AI Usage Control and Agentic AI Security pages and press releases for a published inline decision latency and for fail-open or fail-closed behavior and timeouts; none stated.Source: [sentinelone.com](https://www.sentinelone.com/platform/securing-ai-prompt/), read 2026-11-16. - **sentinelone-52** (vendor): Prompt Security says a skill can redirect a package install to a substituted source, and recommends approved sources only and pinned versions for what agents may install.Quote: "approved sources only, pinned versions, no shortcuts through unofficial channels"Source: [prompt.security](https://prompt.security/blog/when-ai-agents-become-the-supply-chain-hidden-control-planes-in-agentic-systems), read 2026-11-16. - **sentinelone-53** (none-found): None found on 2026-11-16: looked on the Prompt Security pages for blocking of shell commands, file reads or writes, network calls, or credential access by agents on the machine as named stoppable actions; only general runtime policy and sensitive data redaction were found.Source: [sentinelone.com](https://www.sentinelone.com/platform/securing-ai-prompt/), read 2026-11-16. - **sentinelone-54** (vendor): SentinelOne says Prompt Security enforces policy at the same speed agents act, but gives no millisecond figure.Quote: "Enforce policy at the same speed agents act"Source: [sentinelone.com](https://www.sentinelone.com/platform/prompt-security-agentic-ai-security/), read 2026-11-16. ## News mentioning SentinelOne (Prompt Security) - [Prompt Security research on agent approval prompts](/news/prompt-security-agent-approval-prompts/) ## Related pages - [All vendor profiles](/vendors/) - [Pillar Security profile](/vendors/pillar-security/) - [Straiker profile](/vendors/straiker/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # Straiker AI runtime security URL: https://airuntimesecuritytools.com/vendors/straiker/ Updated: 2026-11-16 Straiker pairs named stops with a latency number. It says Defend AI detects or blocks file deletion, remote code execution, exfiltration, uploads, clipboard data, prompt injection and tool misuse, at under 300 ms and 98.1 percent detection accuracy, both vendor claims. Coverage runs from coding agents on endpoints to custom agents and model inputs and outputs, deployed by hook, API, SDK, thin client or gateway as SaaS or on premises. Third-party press confirms funding and revenue growth. Gaps from 16 Nov 2026: fail-open or fail-closed behavior and public documentation are none found, and the product page limits blocking to supported execution points. ## Strength and weakness for AI Runtime Security Tools **Strength:** Names the stopped action types (file deletion, remote code execution, exfiltration, prompt injection) and publishes under-300 ms latency (straiker-08, straiker-23, straiker-42). **Weakness:** Fail-open or fail-closed behavior: none found on 2026-11-16, and the product page says blocking happens at "supported execution points" only (straiker-22, straiker-43). ## Scores and evidence Straiker is joint 4th in the overall ranking at 62.5 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Straiker: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Action types stopped before they complete | 7 of 10 | 30 | The vendor claims blocking of exfiltration, remote code execution, file deletion, uploads and clipboard, prompt injection, tool misuse and toxic output. All of it is vendor-only. | [straiker-07](#fact-straiker-07), [straiker-08](#fact-straiker-08), [straiker-09](#fact-straiker-09), [straiker-40](#fact-straiker-40), [straiker-39](#fact-straiker-39), [straiker-23](#fact-straiker-23) | | Layers covered | 7 of 10 | 20 | The vendor claims coverage of coding agents on endpoints, custom agents, and model inputs and outputs. All of it is vendor-only. | [straiker-07](#fact-straiker-07), [straiker-13](#fact-straiker-13), [straiker-40](#fact-straiker-40), [straiker-22](#fact-straiker-22) | | Latency and failure mode | 5 of 10 | 15 | The vendor claims under 300 ms. Fail-open or fail-closed behavior: none found. | [straiker-42](#fact-straiker-42), [straiker-40](#fact-straiker-40), [straiker-43](#fact-straiker-43) | | Deployment | 7 of 10 | 15 | The vendor says deployment takes minutes with hook, API, SDK or thin-client integration, as SaaS or on-prem. All of it is vendor-claimed. | [straiker-20](#fact-straiker-20), [straiker-28](#fact-straiker-28), [straiker-29](#fact-straiker-29), [straiker-19](#fact-straiker-19) | | Detection and response | 5 of 10 | 5 | The vendor claims an audit trail, tracing, and response actions such as session suspension. Named SIEM, EDR or ticketing integrations: none found. | [straiker-24](#fact-straiker-24), [straiker-44](#fact-straiker-44), [straiker-30](#fact-straiker-30), [straiker-32](#fact-straiker-32) | | Transparency | 3 of 10 | 10 | Research is published. Public docs: none found. The observed marketplace listing has no public price. | [straiker-35](#fact-straiker-35), [straiker-33](#fact-straiker-33), [straiker-34](#fact-straiker-34) | | Proof | 8 of 10 | 5 | Third-party press confirms funding and revenue growth, the vendor names customers, and a Gartner Hype Cycle listing is vendor-cited. | [straiker-03](#fact-straiker-03), [straiker-04](#fact-straiker-04), [straiker-45](#fact-straiker-45), [straiker-36](#fact-straiker-36), [straiker-37](#fact-straiker-37) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | joint 4th | 62.5 | | [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) | 6th | 64.5 | | [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) | 5th | 62.5 | | [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) | 6th | 65.5 | | [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) | 7th | 60.5 | | [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) | joint 5th | 65.5 | ## Product, deployment, price and owner - Product assessed: Defend AI (runtime security and guardrails) with the Agentic Kill Switch - Deployment: SaaS, or on premises for regulated environments; hook, API, SDK, webhook, AI sensor, thin client, browser extension or gateway integration - Time to deploy: Minutes, as the vendor says on its product pages - Price: none found on 2026-11-16 on straiker.ai; AWS Marketplace shows one platform dimension under a contract sized with the vendor - Owner: Straiker Inc. (standalone, venture-backed; no parent company found) - Website: straiker.ai ## Facts and sources - **straiker-01** (vendor): Straiker presents itself as "The Agentic Security Company" and says it is staffed from Mountain View and distributed teams.Quote: "Straiker is built by AI security researchers, security engineers, product builders, and enterprise operators working from Mountain View and distributed teams around the world."Source: [straiker.ai](https://www.straiker.ai/about), read 2026-11-16. - **straiker-02** (vendor): Straiker's leadership page lists Ankur Shah as co-founder and CEO and Sreenath Kurupati as co-founder and CTO.Quote: "Ankur Shah is a cybersecurity leader and entrepreneur with a proven track record of building and scaling innovative security solutions."Source: [straiker.ai](https://www.straiker.ai/about), read 2026-11-16. - **straiker-03** (third-party): SiliconANGLE reported on 2026-06-29 that Straiker Inc., based in Mountain View, raised a $64 million Series A led by Marathon Management Partners with Citi Ventures, Illuminate Ventures and Workday Ventures.Quote: "The Series A round was led by Marathon Management Partners alongside Citi Ventures, Illuminate Ventures and Workday Ventures"Source: [siliconangle.com](https://siliconangle.com/2026/06/29/straiker-lands-64m-defend-enterprise-ai-agents-attack/), read 2026-11-16. - **straiker-04** (third-party): On 2026-08-05 Intesa Sanpaolo group venture arm Neva SGR said it joined the $64 million Series A, bringing total capital raised by Straiker to $85 million.Quote: "The transaction brings the total capital raised by the California-based company to"Source: [group.intesasanpaolo.com](https://group.intesasanpaolo.com/en/newsroom/press-releases/2026/08/neva-sgr-invests-in-straiker--a-leader-in-agentic-ai-security), read 2026-11-16. - **straiker-05** (vendor): Straiker's product line is Discover AI, Ascend AI, Defend AI and Straiker Frontiers, plus an Agentic Kill Switch.Quote: "Discover AI maps every agent, tool, MCP, and Skill in your environment."Source: [straiker.ai](https://www.straiker.ai/platform), read 2026-11-16. - **straiker-06** (vendor): Straiker says Discover AI finds sanctioned and shadow AI across endpoint, browser, gateway, cloud and connected SaaS, including coding, productivity and personal agents and local LLMs.Quote: "Find sanctioned and shadow AI across endpoint, browser, gateway, cloud and connected SaaS sources, including coding, productivity and personal agents and local LLMs/runtimes."Source: [straiker.ai](https://www.straiker.ai/products/discover-ai), read 2026-11-16. - **straiker-07** (vendor): Straiker says it supports Cursor Agent, GitHub Copilot, Claude Code and Windsurf through a thin-client integration and detects and blocks exfiltration, remote code execution and destructive infrastructure actions.Quote: "Straiker currently supports Cursor Agent, GitHub Copilot, Claude Code, and Windsurf, with new coding-agent coverage added regularly."Source: [straiker.ai](https://www.straiker.ai/solution/coding-agents), read 2026-11-16. - **straiker-08** (vendor): Straiker says Defend AI blocks actions such as file deletion and configuration changes by coding agents.Quote: "It can detect and block actions such as file deletion and configuration changes while protecting proprietary code and secrets in development environments."Source: [straiker.ai](https://www.straiker.ai/products/defend-ai), read 2026-11-16. - **straiker-09** (vendor): Straiker says its agentic-browser guardrails block prompt injection and govern forms, uploads, downloads, clipboard and screenshots with allowlists and step-level approvals.Quote: "Govern forms, uploads, downloads, clipboard, and screenshots with destination allowlists and step‑level approvals."Source: [straiker.ai](https://www.straiker.ai/solution/runtime-guardrails-for-agentic-web-browsers), read 2026-11-16. - **straiker-10** (vendor): Straiker says Discover AI uses Browser Guard and endpoint or gateway integrations to enforce employee AI usage policies.Quote: "Browser Guard and endpoint/gateway integrations enforce employee usage policies; activity logs capture user sessions, transcripts and supported tool-invocation events."Source: [straiker.ai](https://www.straiker.ai/products/discover-ai), read 2026-11-16. - **straiker-13** (vendor): Straiker says it covers custom agents built on AWS Bedrock AgentCore, Azure AI Foundry, Copilot Studio, MCP and LangChain, integrated by API, SDK, webhook or gateway.Quote: "Straiker supports custom AI agents built with AWS Bedrock AgentCore, Azure AI Foundry, Azure Copilot Studio, MCP, LangChain, and custom orchestration frameworks."Source: [straiker.ai](https://www.straiker.ai/solution/custom-built-agents), read 2026-11-16. - **straiker-17** (vendor): Straiker says its Agentic Kill Switch can revoke an agent's access to tools, data, APIs and systems.Quote: "Revoke access to the tools, data, APIs, and systems the agent can use."Source: [straiker.ai](https://www.straiker.ai/solution/agentic-kill-switch), read 2026-11-16. - **straiker-19** (vendor): Straiker says Defend AI integrates through hooks, APIs and AI gateways.Quote: "Integrations include hooks, APIs and AI gateways."Source: [straiker.ai](https://www.straiker.ai/products/defend-ai), read 2026-11-16. - **straiker-20** (vendor): Straiker says deployment takes minutes with a single hook-based integration or a one-line install via API, SDK, webhook or AI sensor, with no proxies or firewalls required.Quote: "Deploy in minutes with a single hook-based integration. One-line install via API, SDK, webhook, or AI sensor with no thick clients, proxies, firewalls, or infrastructure changes required."Source: [straiker.ai](https://www.straiker.ai/products/defend-ai), read 2026-11-16. - **straiker-22** (vendor): Straiker says Defend AI uses detection models to detect or block prompt injection, jailbreaks, data exfiltration and unsafe agent actions at supported execution points.Quote: "Specialized detection models evaluate available conversation and tool context to detect or block prompt injection, jailbreaks, data exfiltration and unsafe agent actions at supported execution points."Source: [straiker.ai](https://www.straiker.ai/products/defend-ai), read 2026-11-16. - **straiker-23** (vendor): Straiker says Defend AI detects remote code execution, file access violations, suspicious outbound access, destructive commands and data exfiltration, in Detect or Protect modes.Quote: "Detects remote code execution, file access violations, suspicious outbound access, destructive commands, and data exfiltration"Source: [straiker.ai](https://www.straiker.ai/platform), read 2026-11-16. - **straiker-24** (vendor): Straiker says its Agentic Kill Switch cuts off tool access, suspends the session, isolates or takes an agent offline, and preserves the audit trail.Quote: "Cut off tool access, suspend the session, isolate the agent, or shut it down. Stop one agent or all of them, with a record of what happened."Source: [straiker.ai](https://www.straiker.ai/solution/agentic-kill-switch), read 2026-11-16. - **straiker-28** (vendor): Straiker says coding-agent protection uses a thin client that deploys in minutes next to IDE and terminal workflows.Quote: "Straiker uses a thin-client integration that deploys in minutes and runs alongside existing IDE and terminal workflows."Source: [straiker.ai](https://www.straiker.ai/solution/coding-agents), read 2026-11-16. - **straiker-29** (vendor): Straiker says Defend AI is offered as SaaS or on premises for regulated and sovereign environments.Quote: "SaaS or on premises for regulated and sovereign environments"Source: [straiker.ai](https://www.straiker.ai/platform), read 2026-11-16. - **straiker-30** (vendor): Straiker says Defend AI offers Detect or Protect modes with policies customers write themselves.Quote: "Detect or Protect modes, with policies you write yourself"Source: [straiker.ai](https://www.straiker.ai/platform), read 2026-11-16. - **straiker-32** (none-found): None found on 2026-11-16: named SIEM, EDR, identity provider or ticketing integrations on the partner page; only AI platforms, gateways, clouds and Slack, email or webhook alerts were listed.Source: [straiker.ai](https://www.straiker.ai/partner/technology-and-channel-partners), read 2026-11-16. - **straiker-33** (none-found): None found on 2026-11-16: public product documentation; docs.straiker.ai loads an access-code login page.Source: [docs.straiker.ai](https://docs.straiker.ai/), read 2026-11-16. - **straiker-34** (observed): The AWS Marketplace listing (observed) shows a single "Straiker Platform" pricing dimension billed under a 12-month contract sized with the vendor, not a public price.Quote: "This listing has one pricing dimension: Straiker Platform, billed by units of platform access under a contract."Source: [aws.amazon.com](https://aws.amazon.com/marketplace/pp/prodview-4ey26i3k5epmi), read 2026-11-16. - **straiker-35** (vendor): Straiker publishes research on its site, including a Threat Research report and blog posts dated through 2026-10-21, readable without sign-in.Quote: "Explore Straiker STAR’s latest findings on LLM vulnerabilities, agentic AI threats, and runtime guardrails."Source: [straiker.ai](https://www.straiker.ai/research), read 2026-11-16. - **straiker-36** (vendor): Straiker's customer page quotes security leaders at Omada Health, Coupa, American Express Global Business Travel, EnterpriseDB and Automation Anywhere.Quote: "Straiker helped remove the security blindspot in our B2B AI Application with their unique approach with its Ascend AI product."Source: [straiker.ai](https://www.straiker.ai/customers), read 2026-11-16. - **straiker-37** (vendor): Straiker says it was listed as a Sample Vendor in Agentic AI Security in two Gartner Hype Cycle reports published 2026-10-12.Quote: "Straiker is listed as a Sample Vendor in the Agentic AI Security category of two Gartner Hype Cycle reports published October 12, 2026"Source: [straiker.ai](https://www.straiker.ai/products/defend-ai), read 2026-11-16. - **straiker-38** (vendor): Straiker lists awards including Best Cybersecurity Startup and Best AI Security Testing Platform at The Hacker News Cybersecurity Stars Awards 2026 and the CB Insights AI 100 2026.Quote: "At The Hacker News Cybersecurity Stars Awards 2026, Straiker took two of the program's top honors: Best Cybersecurity Startup and Best AI Security Testing Platform for Ascend AI."Source: [straiker.ai](https://www.straiker.ai/awards-and-recognition), read 2026-11-16. - **straiker-39** (vendor): Straiker says Defend AI for productivity agents blocks indirect prompt injection from emails, documents and calendar invites, cross-connector exfiltration and tool misuse.Quote: "Straiker detects and blocks indirect prompt injection via content (emails, documents, calendar invites), data exfiltration through cross-connector pivots, unauthorized system access, tool misuse"Source: [straiker.ai](https://www.straiker.ai/solution/productivity-agents), read 2026-11-16. - **straiker-40** (vendor): Straiker says Defend AI monitors inputs and outputs and blocks prompt injection, exfiltration, tool misuse, remote code execution and toxic output at subsecond latency.Quote: "Defend AI provides real-time protection against prompt injection, data exfiltration, tool misuse, agent hijacking, remote code execution, resource exhaustion, harmful content generation, and toxic outputs."Source: [straiker.ai](https://www.straiker.ai/solution/custom-built-agents), read 2026-11-16. - **straiker-41** (vendor): Straiker says Defend AI identifies malicious or vulnerable MCP servers using its MCP Threat Database.Quote: "Defend AI identifies malicious, poisoned, or vulnerable MCP servers in real time using Straiker's MCP Threat Database"Source: [straiker.ai](https://www.straiker.ai/products/defend-ai), read 2026-11-16. - **straiker-42** (vendor): Straiker says Defend AI achieves 98.1% detection accuracy at under 300 ms latency.Quote: "delivering 6-21x lower false positive rates than frontier model judges with 98.1% detection accuracy at <300ms latency."Source: [straiker.ai](https://www.straiker.ai/products/defend-ai), read 2026-11-16. - **straiker-43** (none-found): None found on 2026-11-16: published fail-open or fail-closed behavior or timeout settings for Defend AI inline decisions on the product page.Source: [straiker.ai](https://www.straiker.ai/products/defend-ai), read 2026-11-16. - **straiker-44** (vendor): Straiker says Defend AI traces every user, model, tool-chain and agent-to-agent interaction for root-cause analysis, anomaly detection and audit readiness.Quote: "Trace every interaction across the user, model, tool chain, and agent-to-agent calls for root-cause analysis"Source: [straiker.ai](https://www.straiker.ai/products/defend-ai), read 2026-11-16. - **straiker-45** (third-party): On 2026-06-29 SiliconANGLE reported Straiker says run-rate revenue grew more than 15-fold in under a year and that it counts frontier AI labs and Fortune 500 companies as customers.Quote: "It counts frontier AI labs and Fortune 500 companies among its customers."Source: [siliconangle.com](https://siliconangle.com/2026/06/29/straiker-lands-64m-defend-enterprise-ai-agents-attack/), read 2026-11-16. - **straiker-46** (vendor): On 2026-08-11 Straiker announced it is an OpenAI Select Partner, and its partner page says it is an Anthropic Compliance Partner integrated with the Claude Compliance API.Quote: "Straiker is an Anthropic Compliance Partner, integrated with the Claude Compliance API."Source: [straiker.ai](https://www.straiker.ai/partner/technology-and-channel-partners), read 2026-11-16. ## Related pages - [All vendor profiles](/vendors/) - [SentinelOne (Prompt Security) profile](/vendors/sentinelone/) - [Zenity profile](/vendors/zenity/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # Zenity AI runtime security URL: https://airuntimesecuritytools.com/vendors/zenity/ Updated: 2026-11-16 Zenity concentrates on agents and says it stops harm in the middle of a chain. Boundaries stops multi-step exfiltration and privilege-escalation chains before they complete, and native hooks plus an MCP gateway can block or modify a dangerous tool call, including malicious package installs and network calls to listed domains. Its AI detection and response tracks each agent step in one activity table, and detections map to OWASP LLM and MITRE ATLAS. Third-party press confirms its funding and research. Gaps from 16 Nov 2026: inline latency, fail-open or fail-closed behavior, a price and documentation readable without sign-in are none found, since the docs sit behind a login. ## Strength and weakness for AI Runtime Security Tools **Strength:** Documents blocking of tool calls, multi-step exfiltration chains, prompt injection and malicious package installs in real time across agent and endpoint-hook layers (zenity-46, zenity-50, zenity-52, zenity-51). **Weakness:** Published latency and fail-open or fail-closed behavior are none found on 2026-11-16, and docs are login-gated (zenity-56, zenity-39). ## Scores and evidence Zenity is 12th in the overall ranking at 50.0 of 100. Each score below lists the fact ids behind it; the facts follow further down with their sources and read dates. **Zenity: score per criterion** | Criterion | Score | Weight | Why | Facts | |---|---|---|---|---| | Action types stopped before they complete | 7 of 10 | 30 | The vendor says it blocks tool calls, prompt injection, data leakage, secret exfiltration, malicious package installs and network calls to listed domains before execution, all on vendor claims. Blocking of harmful output as a separate control: none found. | [zenity-24](#fact-zenity-24), [zenity-50](#fact-zenity-50), [zenity-51](#fact-zenity-51), [zenity-52](#fact-zenity-52), [zenity-53](#fact-zenity-53), [zenity-48](#fact-zenity-48), [zenity-38](#fact-zenity-38), [zenity-57](#fact-zenity-57) | | Layers covered | 6 of 10 | 20 | The vendor claims enforcement across agent tool calls, AI applications and coding agents on developer machines. Endpoint install details: none found, and all of it is vendor-only. | [zenity-45](#fact-zenity-45), [zenity-47](#fact-zenity-47), [zenity-20](#fact-zenity-20), [zenity-50](#fact-zenity-50), [zenity-29](#fact-zenity-29) | | Latency and failure mode | 0 of 10 | 15 | Published inline latency and fail-open or fail-closed behavior: none found. | [zenity-56](#fact-zenity-56) | | Deployment | 4 of 10 | 15 | The vendor describes API integrations, native hooks and an MCP gateway URL. Operating systems, install method, time to deploy and readable docs: none found. | [zenity-25](#fact-zenity-25), [zenity-26](#fact-zenity-26), [zenity-37](#fact-zenity-37), [zenity-29](#fact-zenity-29), [zenity-58](#fact-zenity-58), [zenity-39](#fact-zenity-39) | | Detection and response | 6 of 10 | 5 | The vendor claims a live view, audit trails, an agent activity table and rule preview, and its docs navigation lists Cribl, CrowdStrike and Palo Alto integrations. Integration detail: none found, because the docs are behind sign-in. | [zenity-27](#fact-zenity-27), [zenity-28](#fact-zenity-28), [zenity-54](#fact-zenity-54), [zenity-49](#fact-zenity-49), [zenity-37](#fact-zenity-37) | | Transparency | 4 of 10 | 10 | Third-party press confirms the vendor's published research. Documentation readable without sign-in and a published price: none found. | [zenity-12](#fact-zenity-12), [zenity-13](#fact-zenity-13), [zenity-14](#fact-zenity-14), [zenity-39](#fact-zenity-39), [zenity-59](#fact-zenity-59) | | Proof | 8 of 10 | 5 | Third-party press confirms the funding, and the vendor names a customer and cites Gartner, Latio and SACR coverage. The analyst coverage is vendor-reported. | [zenity-05](#fact-zenity-05), [zenity-02](#fact-zenity-02), [zenity-04](#fact-zenity-04), [zenity-08](#fact-zenity-08), [zenity-06](#fact-zenity-06) | ## Where it ranks | View | Rank | Total | |---|---|---| | [Overall ranking](/ranking/) | 12th | 50.0 | | [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) | joint 10th | 54.5 | | [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) | 12th | 44.5 | | [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) | 11th | 57.0 | | [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) | 9th | 54.5 | | [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) | 14th | 49.5 | ## Product, deployment, price and owner - Product assessed: Zenity Boundaries (runtime enforcement engine) and AI Detection & Response (AIDR), with native agent hooks and the Zenity MCP gateway for coding agents - Deployment: Connects to platforms by API and integration (Microsoft, Salesforce, OpenAI, Google, AWS, Okta and others per docs navigation); coding agents via native hooks, OpenTelemetry and the Zenity MCP gateway; no agent or proxy install details published - Time to deploy: none found on 2026-11-16 - Price: none found on 2026-11-16 - Owner: Zenity (standalone, venture-backed company; no parent company stated on its pages) - Website: [zenity.io](https://zenity.io) ## Facts and sources - **zenity-01** (vendor): Zenity describes itself as an AI agent security platform that spans discovery, policy, identity, and runtime defense.Quote: "Zenity secures long-horizon AI agents at the decision, the exact point where context and intent come together to create the highest risk."Source: [zenity.io](https://zenity.io/), read 2026-11-16. - **zenity-02** (vendor): On 2026-08-03, Zenity announced a $125 million Series C led by Norwest; it is a standalone venture-backed company headquartered in New York.Quote: "today announced a $125 million Series C led by Norwest."Source: [zenity.io](https://zenity.io/press-release/zenity-raises-125-million-to-secure-the-era-of-1-billion-ai-agents), read 2026-11-16. - **zenity-03** (vendor): Zenity lists its funding history on its company page: a $5M seed in 2021, a $16.5M Series A in 2023 and a $38M Series B in 2024.Quote: "Zenity raises $38M Series B round of fundraising, co-led by Third Point Ventures and DTCP"Source: [zenity.io](https://zenity.io/company), read 2026-11-16. - **zenity-04** (vendor): Zenity says most of its customers are Fortune 500 and Global 2000 organizations and names SoftBank Corp. as a customer.Quote: "including SoftBank Corp."Source: [zenity.io](https://zenity.io/press-release/zenity-raises-125-million-to-secure-the-era-of-1-billion-ai-agents), read 2026-11-16. - **zenity-05** (third-party): Israeli business outlet Calcalist CTech, on 2026-09-03, reports that Zenity recently raised $125 million in a round led by Norwest.Quote: "a company that recently raised $125 million in a funding round led by Norwest"Source: [calcalistech.com](https://www.calcalistech.com/ctechnews/article/rsdllsuy3), read 2026-11-16. - **zenity-06** (vendor): On 2026-09-15, Zenity announced recognition in two independent analyst reports, from Latio and Software Analyst Cyber Research (SACR).Quote: "today announced recognition in two independent analyst reports examining the evolving AI agent security market."Source: [zenity.io](https://zenity.io/press-release/zenity-earns-recognition-in-ai-agent-security-reports-from-latio-and-sacr), read 2026-11-16. - **zenity-07** (vendor): Zenity says SACR places it among vendors combining runtime behavior analysis, MCP and tool governance, agent identity, audit evidence and response.Quote: "SACR identifies Zenity among vendors that combine multiple agent security capabilities, including runtime behavior analysis, MCP and tool governance"Source: [zenity.io](https://zenity.io/press-release/zenity-earns-recognition-in-ai-agent-security-reports-from-latio-and-sacr), read 2026-11-16. - **zenity-08** (vendor): On 2026-09-23, Zenity announced it was named a Market Shaper in Gartner's first Emerging Market Quadrant for AI Application Security (startup vendors).Quote: "Zenity has been named a Market Shaper in Gartner"Source: [zenity.io](https://zenity.io/company/newsroom/press-releases), read 2026-11-16. - **zenity-09** (vendor): Zenity lists a Gartner paper naming it the "Company to Beat" in AI agent governance on its platform pages.Quote: "Gartner® Names Zenity as the Company to Beat in AI Agent Governance"Source: [zenity.io](https://zenity.io/platform/mcp-security), read 2026-11-16. - **zenity-10** (vendor): On 2026-01-06, Zenity announced it is available on AWS Marketplace with support for Amazon Bedrock AgentCore, and says its coverage spans cloud, SaaS, endpoints and agentic browsers.Quote: "across cloud, SaaS, endpoints, and agentic browsers"Source: [zenity.io](https://zenity.io/press-release/zenity-now-available-on-aws-marketplace), read 2026-11-16. - **zenity-12** (third-party): On 2026-09-24, The Register reported that Zenity Labs found three Salesforce Agentforce flaws (SalesBleed) and reported them to Salesforce, which fixed them.Quote: "Zenity Labs uncovered the three vulnerabilities, collectively called SalesBleed, and reported them to Salesforce"Source: [theregister.com](https://www.theregister.com/security/2026/09/24/salesforce-agentforce-vulns-allowed-0-click-crm-data-theft-anonymous-phishing/5298958), read 2026-11-16. - **zenity-13** (third-party): On 2026-09-25, SecurityWeek covered the SalesBleed Agentforce flaws that Zenity Labs reported.Quote: "Three vulnerabilities in Salesforce Agentforce could have allowed attackers to hijack trusted agents for sensitive CRM data exfiltration and phishing, Zenity Labs reports."Source: [securityweek.com](https://www.securityweek.com/salesbleed-flaws-in-salesforce-agentforce-enabled-zero-click-data-exfiltration/), read 2026-11-16. - **zenity-14** (vendor): On 2026-10-08, Zenity Labs published research on AgentCorruption, a chain of AWS AgentCore flaws.Quote: "Zenity Labs Discloses AgentCorruption, a Chain of AWS AgentCore Flaws That Allowed One Prompt to Take Over All AgentCore Agents"Source: [zenity.io](https://zenity.io/company/newsroom/press-releases), read 2026-11-16. - **zenity-20** (vendor): Zenity names Claude Code and Cowork, ChatGPT Codex, GitHub Copilot and Cursor as the coding and personal agents where it enforces boundaries.Quote: "enforces hard boundaries across Anthropic's Claude Code and Cowork, ChatGPT Codex, GitHub Copilot, and Cursor."Source: [zenity.io](https://zenity.io/use-cases/agent-type/coding-personal-agents), read 2026-11-16. - **zenity-24** (vendor): Zenity says its MCP product can allow, modify or block a tool call based on live context.Quote: "Allow, modify, or block a tool call based on live context, not a static allow-list."Source: [zenity.io](https://zenity.io/platform/mcp-security), read 2026-11-16. - **zenity-25** (vendor): Zenity says MCP traffic is controlled by publishing a server and registering one gateway URL, so agents keep working unchanged.Quote: "Publish a server, register one gateway URL, and agents keep working unchanged."Source: [zenity.io](https://zenity.io/platform/mcp-security), read 2026-11-16. - **zenity-26** (vendor): Zenity says it collects coding-agent telemetry through native hooks and OpenTelemetry.Quote: "through native hooks and OpenTelemetry: each step an agent takes, the code and data it touches"Source: [zenity.io](https://zenity.io/use-cases/agent-type/coding-personal-agents), read 2026-11-16. - **zenity-27** (vendor): Zenity says its MCP Live View shows each request, tool invoked and identity behind it in real time.Quote: "Live View shows each request, the tool invoked, and the identity behind it, in real time."Source: [zenity.io](https://zenity.io/platform/mcp-security), read 2026-11-16. - **zenity-28** (vendor): Zenity says it provides unified visibility, audit trails and enforcement across coding and personal agents.Quote: "Prove governance across every coding and personal agent, with unified visibility, audit trails, and enforcement."Source: [zenity.io](https://zenity.io/use-cases/agent-type/coding-personal-agents), read 2026-11-16. - **zenity-29** (none-found): None found on 2026-11-16: operating systems supported, endpoint install method or time to deploy on developer machines, searched on the coding and personal agents page, MCP security page and platform page.Source: [zenity.io](https://zenity.io/use-cases/agent-type/coding-personal-agents), read 2026-11-16. - **zenity-35** (vendor): Zenity says it uses identity to gate agent actions, such as blocking a deactivated Okta user or restricting privileged tools by role.Quote: "Zenity uses identity to gate what an agent can do, like blocking a deactivated Okta user or restricting privileged tools by role."Source: [zenity.io](https://zenity.io/platform/agentic-identity-and-access-management), read 2026-11-16. - **zenity-36** (vendor): Zenity says it flags risky non-human identities such as static API keys that coding agents use and that never appear in the IdP.Quote: "Surface risky NHIs, including static API keys and secrets that coding agents use and that never appear in your IdP."Source: [zenity.io](https://zenity.io/platform/agentic-identity-and-access-management), read 2026-11-16. - **zenity-37** (vendor): Zenity's public docs navigation lists integrations for OpenAI, Google, AWS, Salesforce, Anthropic, Cursor, GitHub, LiteLLM, Cribl Stream, Okta, CrowdStrike and Palo Alto Networks.Quote: "Gateways and custom agents LiteLLM Custom agents Cribl Stream ACS schema Identity and endpoint security Okta CrowdStrike Palo Alto Networks"Source: [docs.zenity.io](https://docs.zenity.io), read 2026-11-16. - **zenity-38** (vendor): Zenity's docs navigation lists pages for Boundaries, Inline Prevention, Kill switch, Malicious Package Protection and AI Detection & Response.Quote: "Protect Incidents AI Detection & Response Malicious Package Protection Enforce Boundaries Boundaries demos Boundaries CLI Inline Prevention Kill switch"Source: [docs.zenity.io](https://docs.zenity.io), read 2026-11-16. - **zenity-39** (none-found): None found on 2026-11-16: documentation body text readable without sign-in; docs.zenity.io/user-guide/* and /release-notes redirected to a login page, and only the navigation tree of the docs home page loaded.Source: [docs.zenity.io](https://docs.zenity.io/user-guide/inline-prevention), read 2026-11-16. - **zenity-42** (vendor): Zenity says inventory appears the moment a platform is connected, with no separate discovery step.Quote: "Full inventory the moment a platform connects, no separate discovery step required"Source: [zenity.io](https://zenity.io/use-cases/risk-type/shadow-ai), read 2026-11-16. - **zenity-45** (vendor): Zenity says its Boundaries engine evaluates every agent action in real time and can let it through, block it, or shut the agent down.Quote: "evaluates every agent action in real time against rules of an organization and decides to let it through, block it, or shut the agent down."Source: [zenity.io](https://zenity.io/platform/runtime-boundaries), read 2026-11-16. - **zenity-46** (vendor): Zenity says Boundaries stops multi-step exfiltration and privilege-escalation chains before they complete.Quote: "Multi-step exfiltration and privilege-escalation chains get stopped before they complete, not flagged after the fact."Source: [zenity.io](https://zenity.io/platform/runtime-boundaries), read 2026-11-16. - **zenity-47** (vendor): Zenity says one rule is enforced across Copilot, ChatGPT Enterprise, Agentforce, Bedrock, coding agents and homegrown builds.Quote: "Write a rule once and it's already enforced on Copilot, ChatGPT Enterprise, Agentforce, Bedrock, coding agents, and homegrown builds"Source: [zenity.io](https://zenity.io/platform/runtime-boundaries), read 2026-11-16. - **zenity-48** (vendor): Zenity gives an example of one rule blocking tool calls to specific third-party domains across every agent.Quote: "A global asset manager blocks tool calls to specific third-party domains across every agent with one rule"Source: [zenity.io](https://zenity.io/platform/runtime-boundaries), read 2026-11-16. - **zenity-49** (vendor): Zenity says a rule can be previewed to show what it would have caught before it can block anything.Quote: "See exactly what a rule would have caught before it can block anything"Source: [zenity.io](https://zenity.io/platform/runtime-boundaries), read 2026-11-16. - **zenity-50** (vendor): Zenity says that through native agent hooks and the Zenity MCP gateway it can block or modify a dangerous tool call before it executes.Quote: "it can block or modify a dangerous tool call before it executes."Source: [zenity.io](https://zenity.io/use-cases/agent-type/coding-personal-agents), read 2026-11-16. - **zenity-51** (vendor): Zenity names secret exfiltration, malicious package installs and unsafe tool chaining as risks it detects and prevents on coding agents.Quote: "Detect and prevent risks like secret exfiltration, malicious package installs, unsafe tool chaining, and an agent working around a security control"Source: [zenity.io](https://zenity.io/use-cases/agent-type/coding-personal-agents), read 2026-11-16. - **zenity-52** (vendor): Zenity says AIDR detects direct and indirect prompt injection and blocks execution before the agent acts on it.Quote: "AIDR detects direct and indirect prompt injection and blocks execution before the agent acts on them."Source: [zenity.io](https://zenity.io/platform/ai-detection-and-response), read 2026-11-16. - **zenity-53** (vendor): Zenity says AIDR monitors and blocks sensitive data leaving through agent conversations and tool calls.Quote: "Real-time monitoring and blocking of sensitive data leakage leaving through agent conversations, tool calls, or encoded payloads"Source: [zenity.io](https://zenity.io/platform/ai-detection-and-response), read 2026-11-16. - **zenity-54** (vendor): Zenity says AIDR tracks each agent step, including triggers, retrievals, tool calls and agent handoffs, in one activity table.Quote: "Step-by-step tracking across triggers, retrievals, tool invocations, and agent handoffs."Source: [zenity.io](https://zenity.io/platform/ai-detection-and-response), read 2026-11-16. - **zenity-55** (vendor): Zenity says AIDR detections map to OWASP LLM and MITRE ATLAS and can block unsafe actions in real time.Quote: "detecting threats mapped to the OWASP LLM and MITRE ATLAS frameworks and blocking unsafe actions in real time."Source: [zenity.io](https://zenity.io/platform/ai-detection-and-response), read 2026-11-16. - **zenity-56** (none-found): None found on 2026-11-16: published latency of an inline decision, or published fail-open or fail-closed behavior and timeouts, searched on the Boundaries, AIDR, MCP and coding agent pages.Source: [zenity.io](https://zenity.io/platform/runtime-boundaries), read 2026-11-16. - **zenity-57** (none-found): None found on 2026-11-16: documented blocking of harmful or toxic model output as a separate control, searched on the AIDR and Boundaries pages (data leakage and prompt injection are documented).Source: [zenity.io](https://zenity.io/platform/ai-detection-and-response), read 2026-11-16. - **zenity-58** (none-found): None found on 2026-11-16: a published time to deploy for the platform, searched on the platform page, product pages and press releases.Source: [zenity.io](https://zenity.io/platform), read 2026-11-16. - **zenity-59** (none-found): None found on 2026-11-16: a published price or pricing page; https://zenity.io/pricing returned HTTP 404 and the sitemap lists no pricing page.Source: [zenity.io](https://zenity.io/pricing), read 2026-11-16. ## News mentioning Zenity - [AgentCorruption: AWS makes IMDSv2 default in AgentCore](/news/zenity-agentcorruption-aws-agentcore-fix/) ## Related pages - [All vendor profiles](/vendors/) - [Straiker profile](/vendors/straiker/) - [Aim Security profile](/vendors/aim-security/) - [The ranking](/ranking/) - [How scores are set](/method/) - [Best tool by situation](/best-for/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # Best AI runtime security tool by situation URL: https://airuntimesecuritytools.com/best-for/ Updated: 2026-11-16 Five situation pages re-weight the same 0 to 10 scores for coding agents, customer-facing LLM apps, SaaS agents, regulated response and fast rollout. Weights were fixed on 16 Nov 2026 before scoring, so the leader changes only when the question does. ## 5 situations - [Developers run coding agents with shell, file and package access](/best-for/coding-agents/): HiddenLayer leads at 74.5; Bay is 3rd. - [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/): HiddenLayer leads at 68.0; Bay is joint 3rd. - [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/): HiddenLayer leads at 77.5; Bay is joint 4th. - [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/): HiddenLayer leads at 77.0; Bay is joint 3rd. - [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/): HiddenLayer leads at 69.5; Bay is 2nd. ## Who should pick what With no particular situation, the published ranking applies and HiddenLayer comes first. For a specific situation the weights change, and so can the order: - **[Developers run coding agents with shell, file and package access](/best-for/coding-agents/)**: pick HiddenLayer (74.5). Bay is 3rd at 67.5. A coding agent acts on the machine, so stopping named actions there before they complete weighs most. - **[A company shipping a customer-facing LLM application](/best-for/customer-llm-app/)**: pick HiddenLayer (68.0). Bay is joint 3rd at 63.5. An inline control in a customer path must be fast and fail predictably, so latency and failure mode weigh more. - **[Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/)**: pick HiddenLayer (77.5). Bay is joint 4th at 67.0. Agents in SaaS act far from the endpoint, so coverage across layers weighs as much as what is stopped. - **[A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/)**: pick HiddenLayer (77.0). Bay is joint 3rd at 66.5. Proving control means an audit trail and a response path, so detection and response weighs most after what is stopped. - **[A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/)**: pick HiddenLayer (69.5). Bay is 2nd at 69.0. A lean team with a deadline needs the lightest deployment that still stops actions. ## Related pages - [The overall ranking](/ranking/) - [Weights per situation](/method/) - [Vendor profiles](/vendors/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # Which tool is best when developers run coding agents with shell, file and package access? URL: https://airuntimesecuritytools.com/best-for/coding-agents/ Updated: 2026-11-16 HiddenLayer leads when developers run coding agents with shell, file and package access, on weights fixed 16 Nov 2026. Stops weighs 35 and detection and response 10. A tool must block shell commands, file reads and package installs on the machine before they complete, and show what it blocked. ## The short answer HiddenLayer leads the view "Developers run coding agents with shell, file and package access" at 74.5 of 100. Bay is 3rd at 67.5. HiddenLayer is ahead mainly on layers covered (9 against 6, worth 6.0 points at weight 20). ## Why these weights A coding agent acts on the machine, so stopping named actions there before they complete weighs most. **Weights used in this view** | Action types stopped before they complete | Layers covered | Latency and failure mode | Deployment | Detection and response | Transparency | Proof | |---|---|---|---|---|---|---| | 35 | 20 | 10 | 15 | 10 | 5 | 5 | **Developers run coding agents with shell, file and package access: scores out of 10 per criterion, totals out of 100** | Rank | Vendor | Total (of 100) | Action types stopped before they complete (weight 35) | Layers covered (weight 20) | Latency and failure mode (weight 10) | Deployment (weight 15) | Detection and response (weight 10) | Transparency (weight 5) | Proof (weight 5) | |---|---|---|---|---|---|---|---|---|---| | 1st | [HiddenLayer](/vendors/hiddenlayer/) | 74.5 | [8](/vendors/hiddenlayer/#evidence-stops) | [9](/vendors/hiddenlayer/#evidence-layers) | [4](/vendors/hiddenlayer/#evidence-latency) | [6](/vendors/hiddenlayer/#evidence-deploy) | [9](/vendors/hiddenlayer/#evidence-response) | [6](/vendors/hiddenlayer/#evidence-transparency) | [7](/vendors/hiddenlayer/#evidence-proof) | | 2nd | [Cisco](/vendors/cisco/) | 69.0 | [7](/vendors/cisco/#evidence-stops) | [8](/vendors/cisco/#evidence-layers) | [5](/vendors/cisco/#evidence-latency) | [7](/vendors/cisco/#evidence-deploy) | [8](/vendors/cisco/#evidence-response) | [5](/vendors/cisco/#evidence-transparency) | [5](/vendors/cisco/#evidence-proof) | | 3rd | [Bay](/vendors/bay/) | 67.5 | [8](/vendors/bay/#evidence-stops) | [6](/vendors/bay/#evidence-layers) | [5](/vendors/bay/#evidence-latency) | [8](/vendors/bay/#evidence-deploy) | [7](/vendors/bay/#evidence-response) | [0](/vendors/bay/#evidence-transparency) | [7](/vendors/bay/#evidence-proof) | | 4th | [Lasso Security](/vendors/lasso-security/) | 66.5 | [7](/vendors/lasso-security/#evidence-stops) | [7](/vendors/lasso-security/#evidence-layers) | [5](/vendors/lasso-security/#evidence-latency) | [7](/vendors/lasso-security/#evidence-deploy) | [7](/vendors/lasso-security/#evidence-response) | [3](/vendors/lasso-security/#evidence-transparency) | [8](/vendors/lasso-security/#evidence-proof) | | 5th | [Operant AI](/vendors/operant-ai/) | 65.5 | [7](/vendors/operant-ai/#evidence-stops) | [7](/vendors/operant-ai/#evidence-layers) | [4](/vendors/operant-ai/#evidence-latency) | [7](/vendors/operant-ai/#evidence-deploy) | [7](/vendors/operant-ai/#evidence-response) | [4](/vendors/operant-ai/#evidence-transparency) | [7](/vendors/operant-ai/#evidence-proof) | | 6th | [Straiker](/vendors/straiker/) | 64.5 | [7](/vendors/straiker/#evidence-stops) | [7](/vendors/straiker/#evidence-layers) | [5](/vendors/straiker/#evidence-latency) | [7](/vendors/straiker/#evidence-deploy) | [5](/vendors/straiker/#evidence-response) | [3](/vendors/straiker/#evidence-transparency) | [8](/vendors/straiker/#evidence-proof) | | 7th | [Pillar Security](/vendors/pillar-security/) | 60.5 | [7](/vendors/pillar-security/#evidence-stops) | [7](/vendors/pillar-security/#evidence-layers) | [0](/vendors/pillar-security/#evidence-latency) | [6](/vendors/pillar-security/#evidence-deploy) | [7](/vendors/pillar-security/#evidence-response) | [4](/vendors/pillar-security/#evidence-transparency) | [8](/vendors/pillar-security/#evidence-proof) | | 8th | [Palo Alto Networks](/vendors/palo-alto-networks/) | 58.5 | [6](/vendors/palo-alto-networks/#evidence-stops) | [7](/vendors/palo-alto-networks/#evidence-layers) | [4](/vendors/palo-alto-networks/#evidence-latency) | [6](/vendors/palo-alto-networks/#evidence-deploy) | [4](/vendors/palo-alto-networks/#evidence-response) | [6](/vendors/palo-alto-networks/#evidence-transparency) | [7](/vendors/palo-alto-networks/#evidence-proof) | | 9th | [CrowdStrike](/vendors/crowdstrike/) | 56.5 | [6](/vendors/crowdstrike/#evidence-stops) | [7](/vendors/crowdstrike/#evidence-layers) | [0](/vendors/crowdstrike/#evidence-latency) | [6](/vendors/crowdstrike/#evidence-deploy) | [7](/vendors/crowdstrike/#evidence-response) | [4](/vendors/crowdstrike/#evidence-transparency) | [7](/vendors/crowdstrike/#evidence-proof) | | joint 10th | [Aim Security](/vendors/aim-security/) | 54.5 | [6](/vendors/aim-security/#evidence-stops) | [7](/vendors/aim-security/#evidence-layers) | [0](/vendors/aim-security/#evidence-latency) | [6](/vendors/aim-security/#evidence-deploy) | [5](/vendors/aim-security/#evidence-response) | [3](/vendors/aim-security/#evidence-transparency) | [8](/vendors/aim-security/#evidence-proof) | | joint 10th | [Zenity](/vendors/zenity/) | 54.5 | [7](/vendors/zenity/#evidence-stops) | [6](/vendors/zenity/#evidence-layers) | [0](/vendors/zenity/#evidence-latency) | [4](/vendors/zenity/#evidence-deploy) | [6](/vendors/zenity/#evidence-response) | [4](/vendors/zenity/#evidence-transparency) | [8](/vendors/zenity/#evidence-proof) | | 12th | [Check Point](/vendors/check-point/) | 54.0 | [5](/vendors/check-point/#evidence-stops) | [6](/vendors/check-point/#evidence-layers) | [5](/vendors/check-point/#evidence-latency) | [6](/vendors/check-point/#evidence-deploy) | [4](/vendors/check-point/#evidence-response) | [6](/vendors/check-point/#evidence-transparency) | [7](/vendors/check-point/#evidence-proof) | | 13th | [Cloudflare](/vendors/cloudflare/) | 48.5 | [5](/vendors/cloudflare/#evidence-stops) | [5](/vendors/cloudflare/#evidence-layers) | [0](/vendors/cloudflare/#evidence-latency) | [6](/vendors/cloudflare/#evidence-deploy) | [6](/vendors/cloudflare/#evidence-response) | [6](/vendors/cloudflare/#evidence-transparency) | [6](/vendors/cloudflare/#evidence-proof) | | 14th | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | 48.0 | [5](/vendors/sentinelone/#evidence-stops) | [6](/vendors/sentinelone/#evidence-layers) | [0](/vendors/sentinelone/#evidence-latency) | [6](/vendors/sentinelone/#evidence-deploy) | [4](/vendors/sentinelone/#evidence-response) | [4](/vendors/sentinelone/#evidence-transparency) | [7](/vendors/sentinelone/#evidence-proof) | | 15th | [Microsoft](/vendors/microsoft/) | 47.0 | [4](/vendors/microsoft/#evidence-stops) | [6](/vendors/microsoft/#evidence-layers) | [0](/vendors/microsoft/#evidence-latency) | [6](/vendors/microsoft/#evidence-deploy) | [6](/vendors/microsoft/#evidence-response) | [6](/vendors/microsoft/#evidence-transparency) | [6](/vendors/microsoft/#evidence-proof) | | 16th | [Noma Security](/vendors/noma-security/) | 44.5 | [5](/vendors/noma-security/#evidence-stops) | [6](/vendors/noma-security/#evidence-layers) | [0](/vendors/noma-security/#evidence-latency) | [4](/vendors/noma-security/#evidence-deploy) | [4](/vendors/noma-security/#evidence-response) | [3](/vendors/noma-security/#evidence-transparency) | [7](/vendors/noma-security/#evidence-proof) | ## What would change this ranking **Top pair.** Cisco rising from 7 to 9 on action types stopped before they complete (weight 35, so each point moves a total by 3.5) would put Cisco above HiddenLayer; the gap is 5.5 points. **Bay's next place.** Bay rising from 8 to 9 on action types stopped before they complete (weight 35, so each point moves a total by 3.5) would put Bay above Cisco; the gap is 1.5 points. A score changes only when a sourced fact changes, and the weights were fixed before the scores. The [method page](/method/) lists the weights, anchors and rules. ## Which tool is best when developers run coding agents with shell, file and package access? HiddenLayer leads this view. A coding agent acts on the developer machine, so the weighting gives 35 of 100 to the actions a tool stops before they complete and 20 to the layers it covers, which includes the endpoint. Check three things in the vendor evidence before choosing: that shell command, file access and package install are each named, that the block happens before the action completes, and that the tool supports the agents your developers use. Package install blocking is the thinnest area across the field, with none found for most vendors on 16 Nov 2026. ## What is weighted differently here? Stops 35, layers 20, deployment 15, latency 10, detection and response 10, transparency 5 and proof 5. Latency drops to 10 because a decision on a local hook adds little delay a developer notices, while detection and response rises to 10 so blocked actions leave a trail. ## Related pages - [All situations](/best-for/) - [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) - [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) - [Overall ranking](/ranking/) - [Method and weights](/method/) - [Vendor profiles](/vendors/) --- # Which tool is best when a company ships a customer-facing LLM application? URL: https://airuntimesecuritytools.com/best-for/customer-llm-app/ Updated: 2026-11-16 HiddenLayer leads for a company shipping a customer-facing LLM application, on weights fixed 16 Nov 2026. Latency and failure mode weighs 25, up from 15, because an inline control sits in every customer request and must be fast and predictable when it fails. ## The short answer HiddenLayer leads the view "A company shipping a customer-facing LLM application" at 68.0 of 100. Bay is joint 3rd at 63.5. HiddenLayer is ahead mainly on layers covered (9 against 6, worth 6.0 points at weight 20). ## Why these weights An inline control in a customer path must be fast and fail predictably, so latency and failure mode weigh more. **Weights used in this view** | Action types stopped before they complete | Layers covered | Latency and failure mode | Deployment | Detection and response | Transparency | Proof | |---|---|---|---|---|---|---| | 25 | 20 | 25 | 15 | 5 | 5 | 5 | **A company shipping a customer-facing LLM application: scores out of 10 per criterion, totals out of 100** | Rank | Vendor | Total (of 100) | Action types stopped before they complete (weight 25) | Layers covered (weight 20) | Latency and failure mode (weight 25) | Deployment (weight 15) | Detection and response (weight 5) | Transparency (weight 5) | Proof (weight 5) | |---|---|---|---|---|---|---|---|---|---| | 1st | [HiddenLayer](/vendors/hiddenlayer/) | 68.0 | [8](/vendors/hiddenlayer/#evidence-stops) | [9](/vendors/hiddenlayer/#evidence-layers) | [4](/vendors/hiddenlayer/#evidence-latency) | [6](/vendors/hiddenlayer/#evidence-deploy) | [9](/vendors/hiddenlayer/#evidence-response) | [6](/vendors/hiddenlayer/#evidence-transparency) | [7](/vendors/hiddenlayer/#evidence-proof) | | 2nd | [Cisco](/vendors/cisco/) | 65.5 | [7](/vendors/cisco/#evidence-stops) | [8](/vendors/cisco/#evidence-layers) | [5](/vendors/cisco/#evidence-latency) | [7](/vendors/cisco/#evidence-deploy) | [8](/vendors/cisco/#evidence-response) | [5](/vendors/cisco/#evidence-transparency) | [5](/vendors/cisco/#evidence-proof) | | joint 3rd | [Bay](/vendors/bay/) | 63.5 | [8](/vendors/bay/#evidence-stops) | [6](/vendors/bay/#evidence-layers) | [5](/vendors/bay/#evidence-latency) | [8](/vendors/bay/#evidence-deploy) | [7](/vendors/bay/#evidence-response) | [0](/vendors/bay/#evidence-transparency) | [7](/vendors/bay/#evidence-proof) | | joint 3rd | [Lasso Security](/vendors/lasso-security/) | 63.5 | [7](/vendors/lasso-security/#evidence-stops) | [7](/vendors/lasso-security/#evidence-layers) | [5](/vendors/lasso-security/#evidence-latency) | [7](/vendors/lasso-security/#evidence-deploy) | [7](/vendors/lasso-security/#evidence-response) | [3](/vendors/lasso-security/#evidence-transparency) | [8](/vendors/lasso-security/#evidence-proof) | | 5th | [Straiker](/vendors/straiker/) | 62.5 | [7](/vendors/straiker/#evidence-stops) | [7](/vendors/straiker/#evidence-layers) | [5](/vendors/straiker/#evidence-latency) | [7](/vendors/straiker/#evidence-deploy) | [5](/vendors/straiker/#evidence-response) | [3](/vendors/straiker/#evidence-transparency) | [8](/vendors/straiker/#evidence-proof) | | 6th | [Operant AI](/vendors/operant-ai/) | 61.0 | [7](/vendors/operant-ai/#evidence-stops) | [7](/vendors/operant-ai/#evidence-layers) | [4](/vendors/operant-ai/#evidence-latency) | [7](/vendors/operant-ai/#evidence-deploy) | [7](/vendors/operant-ai/#evidence-response) | [4](/vendors/operant-ai/#evidence-transparency) | [7](/vendors/operant-ai/#evidence-proof) | | 7th | [Palo Alto Networks](/vendors/palo-alto-networks/) | 56.5 | [6](/vendors/palo-alto-networks/#evidence-stops) | [7](/vendors/palo-alto-networks/#evidence-layers) | [4](/vendors/palo-alto-networks/#evidence-latency) | [6](/vendors/palo-alto-networks/#evidence-deploy) | [4](/vendors/palo-alto-networks/#evidence-response) | [6](/vendors/palo-alto-networks/#evidence-transparency) | [7](/vendors/palo-alto-networks/#evidence-proof) | | 8th | [Check Point](/vendors/check-point/) | 54.5 | [5](/vendors/check-point/#evidence-stops) | [6](/vendors/check-point/#evidence-layers) | [5](/vendors/check-point/#evidence-latency) | [6](/vendors/check-point/#evidence-deploy) | [4](/vendors/check-point/#evidence-response) | [6](/vendors/check-point/#evidence-transparency) | [7](/vendors/check-point/#evidence-proof) | | 9th | [Pillar Security](/vendors/pillar-security/) | 50.0 | [7](/vendors/pillar-security/#evidence-stops) | [7](/vendors/pillar-security/#evidence-layers) | [0](/vendors/pillar-security/#evidence-latency) | [6](/vendors/pillar-security/#evidence-deploy) | [7](/vendors/pillar-security/#evidence-response) | [4](/vendors/pillar-security/#evidence-transparency) | [8](/vendors/pillar-security/#evidence-proof) | | 10th | [CrowdStrike](/vendors/crowdstrike/) | 47.0 | [6](/vendors/crowdstrike/#evidence-stops) | [7](/vendors/crowdstrike/#evidence-layers) | [0](/vendors/crowdstrike/#evidence-latency) | [6](/vendors/crowdstrike/#evidence-deploy) | [7](/vendors/crowdstrike/#evidence-response) | [4](/vendors/crowdstrike/#evidence-transparency) | [7](/vendors/crowdstrike/#evidence-proof) | | 11th | [Aim Security](/vendors/aim-security/) | 46.0 | [6](/vendors/aim-security/#evidence-stops) | [7](/vendors/aim-security/#evidence-layers) | [0](/vendors/aim-security/#evidence-latency) | [6](/vendors/aim-security/#evidence-deploy) | [5](/vendors/aim-security/#evidence-response) | [3](/vendors/aim-security/#evidence-transparency) | [8](/vendors/aim-security/#evidence-proof) | | 12th | [Zenity](/vendors/zenity/) | 44.5 | [7](/vendors/zenity/#evidence-stops) | [6](/vendors/zenity/#evidence-layers) | [0](/vendors/zenity/#evidence-latency) | [4](/vendors/zenity/#evidence-deploy) | [6](/vendors/zenity/#evidence-response) | [4](/vendors/zenity/#evidence-transparency) | [8](/vendors/zenity/#evidence-proof) | | 13th | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | 41.0 | [5](/vendors/sentinelone/#evidence-stops) | [6](/vendors/sentinelone/#evidence-layers) | [0](/vendors/sentinelone/#evidence-latency) | [6](/vendors/sentinelone/#evidence-deploy) | [4](/vendors/sentinelone/#evidence-response) | [4](/vendors/sentinelone/#evidence-transparency) | [7](/vendors/sentinelone/#evidence-proof) | | 14th | [Cloudflare](/vendors/cloudflare/) | 40.5 | [5](/vendors/cloudflare/#evidence-stops) | [5](/vendors/cloudflare/#evidence-layers) | [0](/vendors/cloudflare/#evidence-latency) | [6](/vendors/cloudflare/#evidence-deploy) | [6](/vendors/cloudflare/#evidence-response) | [6](/vendors/cloudflare/#evidence-transparency) | [6](/vendors/cloudflare/#evidence-proof) | | 15th | [Microsoft](/vendors/microsoft/) | 40.0 | [4](/vendors/microsoft/#evidence-stops) | [6](/vendors/microsoft/#evidence-layers) | [0](/vendors/microsoft/#evidence-latency) | [6](/vendors/microsoft/#evidence-deploy) | [6](/vendors/microsoft/#evidence-response) | [6](/vendors/microsoft/#evidence-transparency) | [6](/vendors/microsoft/#evidence-proof) | | 16th | [Noma Security](/vendors/noma-security/) | 37.5 | [5](/vendors/noma-security/#evidence-stops) | [6](/vendors/noma-security/#evidence-layers) | [0](/vendors/noma-security/#evidence-latency) | [4](/vendors/noma-security/#evidence-deploy) | [4](/vendors/noma-security/#evidence-response) | [3](/vendors/noma-security/#evidence-transparency) | [7](/vendors/noma-security/#evidence-proof) | ## What would change this ranking **Top pair.** Cisco rising from 7 to 9 on action types stopped before they complete (weight 25, so each point moves a total by 2.5) would put Cisco above HiddenLayer; the gap is 2.5 points. **Bay's next place.** Bay rising from 8 to 9 on action types stopped before they complete (weight 25, so each point moves a total by 2.5) would put Bay above Cisco; the gap is 2.0 points. A score changes only when a sourced fact changes, and the weights were fixed before the scores. The [method page](/method/) lists the weights, anchors and rules. ## Which tool is best when a company ships a customer-facing LLM application? HiddenLayer leads this view. In a customer path, a slow or silently failing control is a product risk, so latency and failure mode carries 25 of 100, stops 25, layers 20 and deployment 15. Published numbers are scarce. Check whether the vendor gives a millisecond figure with a stated method, and whether the default on failure is to block or to let traffic through. A vendor claim without a failure mode earns a low latency score by design. ## What is weighted differently here? Latency and failure mode rises from 15 to 25, stops falls from 30 to 25, and transparency falls to 5. Application and model layer enforcement matters more than endpoint coverage for this case, which the layers criterion reflects. ## Related pages - [All situations](/best-for/) - [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) - [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) - [Overall ranking](/ranking/) - [Method and weights](/method/) - [Vendor profiles](/vendors/) --- # Which tool is best when staff use agents inside SaaS and cloud workspaces? URL: https://airuntimesecuritytools.com/best-for/saas-agents/ Updated: 2026-11-16 HiddenLayer leads when staff use agents inside SaaS and cloud workspaces that act on company data, on weights fixed 16 Nov 2026. Layers covered rises to 30, equal to stops, because these agents act far from the endpoint and no single enforcement point sees them. ## The short answer HiddenLayer leads the view "Staff use agents inside SaaS and cloud workspaces that act on company data" at 77.5 of 100. Bay is joint 4th at 67.0. HiddenLayer is ahead mainly on layers covered (9 against 6, worth 9.0 points at weight 30). ## Why these weights Agents in SaaS act far from the endpoint, so coverage across layers weighs as much as what is stopped. **Weights used in this view** | Action types stopped before they complete | Layers covered | Latency and failure mode | Deployment | Detection and response | Transparency | Proof | |---|---|---|---|---|---|---| | 30 | 30 | 5 | 15 | 10 | 5 | 5 | **Staff use agents inside SaaS and cloud workspaces that act on company data: scores out of 10 per criterion, totals out of 100** | Rank | Vendor | Total (of 100) | Action types stopped before they complete (weight 30) | Layers covered (weight 30) | Latency and failure mode (weight 5) | Deployment (weight 15) | Detection and response (weight 10) | Transparency (weight 5) | Proof (weight 5) | |---|---|---|---|---|---|---|---|---|---| | 1st | [HiddenLayer](/vendors/hiddenlayer/) | 77.5 | [8](/vendors/hiddenlayer/#evidence-stops) | [9](/vendors/hiddenlayer/#evidence-layers) | [4](/vendors/hiddenlayer/#evidence-latency) | [6](/vendors/hiddenlayer/#evidence-deploy) | [9](/vendors/hiddenlayer/#evidence-response) | [6](/vendors/hiddenlayer/#evidence-transparency) | [7](/vendors/hiddenlayer/#evidence-proof) | | 2nd | [Cisco](/vendors/cisco/) | 71.0 | [7](/vendors/cisco/#evidence-stops) | [8](/vendors/cisco/#evidence-layers) | [5](/vendors/cisco/#evidence-latency) | [7](/vendors/cisco/#evidence-deploy) | [8](/vendors/cisco/#evidence-response) | [5](/vendors/cisco/#evidence-transparency) | [5](/vendors/cisco/#evidence-proof) | | 3rd | [Lasso Security](/vendors/lasso-security/) | 67.5 | [7](/vendors/lasso-security/#evidence-stops) | [7](/vendors/lasso-security/#evidence-layers) | [5](/vendors/lasso-security/#evidence-latency) | [7](/vendors/lasso-security/#evidence-deploy) | [7](/vendors/lasso-security/#evidence-response) | [3](/vendors/lasso-security/#evidence-transparency) | [8](/vendors/lasso-security/#evidence-proof) | | joint 4th | [Bay](/vendors/bay/) | 67.0 | [8](/vendors/bay/#evidence-stops) | [6](/vendors/bay/#evidence-layers) | [5](/vendors/bay/#evidence-latency) | [8](/vendors/bay/#evidence-deploy) | [7](/vendors/bay/#evidence-response) | [0](/vendors/bay/#evidence-transparency) | [7](/vendors/bay/#evidence-proof) | | joint 4th | [Operant AI](/vendors/operant-ai/) | 67.0 | [7](/vendors/operant-ai/#evidence-stops) | [7](/vendors/operant-ai/#evidence-layers) | [4](/vendors/operant-ai/#evidence-latency) | [7](/vendors/operant-ai/#evidence-deploy) | [7](/vendors/operant-ai/#evidence-response) | [4](/vendors/operant-ai/#evidence-transparency) | [7](/vendors/operant-ai/#evidence-proof) | | 6th | [Straiker](/vendors/straiker/) | 65.5 | [7](/vendors/straiker/#evidence-stops) | [7](/vendors/straiker/#evidence-layers) | [5](/vendors/straiker/#evidence-latency) | [7](/vendors/straiker/#evidence-deploy) | [5](/vendors/straiker/#evidence-response) | [3](/vendors/straiker/#evidence-transparency) | [8](/vendors/straiker/#evidence-proof) | | 7th | [Pillar Security](/vendors/pillar-security/) | 64.0 | [7](/vendors/pillar-security/#evidence-stops) | [7](/vendors/pillar-security/#evidence-layers) | [0](/vendors/pillar-security/#evidence-latency) | [6](/vendors/pillar-security/#evidence-deploy) | [7](/vendors/pillar-security/#evidence-response) | [4](/vendors/pillar-security/#evidence-transparency) | [8](/vendors/pillar-security/#evidence-proof) | | joint 8th | [CrowdStrike](/vendors/crowdstrike/) | 60.5 | [6](/vendors/crowdstrike/#evidence-stops) | [7](/vendors/crowdstrike/#evidence-layers) | [0](/vendors/crowdstrike/#evidence-latency) | [6](/vendors/crowdstrike/#evidence-deploy) | [7](/vendors/crowdstrike/#evidence-response) | [4](/vendors/crowdstrike/#evidence-transparency) | [7](/vendors/crowdstrike/#evidence-proof) | | joint 8th | [Palo Alto Networks](/vendors/palo-alto-networks/) | 60.5 | [6](/vendors/palo-alto-networks/#evidence-stops) | [7](/vendors/palo-alto-networks/#evidence-layers) | [4](/vendors/palo-alto-networks/#evidence-latency) | [6](/vendors/palo-alto-networks/#evidence-deploy) | [4](/vendors/palo-alto-networks/#evidence-response) | [6](/vendors/palo-alto-networks/#evidence-transparency) | [7](/vendors/palo-alto-networks/#evidence-proof) | | 10th | [Aim Security](/vendors/aim-security/) | 58.5 | [6](/vendors/aim-security/#evidence-stops) | [7](/vendors/aim-security/#evidence-layers) | [0](/vendors/aim-security/#evidence-latency) | [6](/vendors/aim-security/#evidence-deploy) | [5](/vendors/aim-security/#evidence-response) | [3](/vendors/aim-security/#evidence-transparency) | [8](/vendors/aim-security/#evidence-proof) | | 11th | [Zenity](/vendors/zenity/) | 57.0 | [7](/vendors/zenity/#evidence-stops) | [6](/vendors/zenity/#evidence-layers) | [0](/vendors/zenity/#evidence-latency) | [4](/vendors/zenity/#evidence-deploy) | [6](/vendors/zenity/#evidence-response) | [4](/vendors/zenity/#evidence-transparency) | [8](/vendors/zenity/#evidence-proof) | | 12th | [Check Point](/vendors/check-point/) | 55.0 | [5](/vendors/check-point/#evidence-stops) | [6](/vendors/check-point/#evidence-layers) | [5](/vendors/check-point/#evidence-latency) | [6](/vendors/check-point/#evidence-deploy) | [4](/vendors/check-point/#evidence-response) | [6](/vendors/check-point/#evidence-transparency) | [7](/vendors/check-point/#evidence-proof) | | 13th | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | 51.5 | [5](/vendors/sentinelone/#evidence-stops) | [6](/vendors/sentinelone/#evidence-layers) | [0](/vendors/sentinelone/#evidence-latency) | [6](/vendors/sentinelone/#evidence-deploy) | [4](/vendors/sentinelone/#evidence-response) | [4](/vendors/sentinelone/#evidence-transparency) | [7](/vendors/sentinelone/#evidence-proof) | | joint 14th | [Cloudflare](/vendors/cloudflare/) | 51.0 | [5](/vendors/cloudflare/#evidence-stops) | [5](/vendors/cloudflare/#evidence-layers) | [0](/vendors/cloudflare/#evidence-latency) | [6](/vendors/cloudflare/#evidence-deploy) | [6](/vendors/cloudflare/#evidence-response) | [6](/vendors/cloudflare/#evidence-transparency) | [6](/vendors/cloudflare/#evidence-proof) | | joint 14th | [Microsoft](/vendors/microsoft/) | 51.0 | [4](/vendors/microsoft/#evidence-stops) | [6](/vendors/microsoft/#evidence-layers) | [0](/vendors/microsoft/#evidence-latency) | [6](/vendors/microsoft/#evidence-deploy) | [6](/vendors/microsoft/#evidence-response) | [6](/vendors/microsoft/#evidence-transparency) | [6](/vendors/microsoft/#evidence-proof) | | 16th | [Noma Security](/vendors/noma-security/) | 48.0 | [5](/vendors/noma-security/#evidence-stops) | [6](/vendors/noma-security/#evidence-layers) | [0](/vendors/noma-security/#evidence-latency) | [4](/vendors/noma-security/#evidence-deploy) | [4](/vendors/noma-security/#evidence-response) | [3](/vendors/noma-security/#evidence-transparency) | [7](/vendors/noma-security/#evidence-proof) | ## What would change this ranking **Top pair.** Cisco rising from 7 to 10 on action types stopped before they complete (weight 30, so each point moves a total by 3.0) would put Cisco above HiddenLayer; the gap is 6.5 points. **Bay's next place.** Bay rising from 8 to 9 on action types stopped before they complete (weight 30, so each point moves a total by 3.0) would put Bay above Lasso Security; the gap is 0.5 points. A score changes only when a sourced fact changes, and the weights were fixed before the scores. The [method page](/method/) lists the weights, anchors and rules. ## Which tool is best when staff use agents inside SaaS and cloud workspaces? HiddenLayer leads this view. Agents in SaaS run on the platform provider's infrastructure, so a control at the endpoint may never see them. Coverage across model, application and agent layers weighs as much as what is stopped. Ask each vendor which agent platforms it integrates with by name, and whether the integration blocks or only detects. Several vendors document a detect-only path for some platforms, which the evidence files record as partial. ## What is weighted differently here? Stops 30, layers 30, deployment 15, detection and response 10, latency 5, transparency 5 and proof 5. Latency falls to 5 because agent actions in SaaS are asynchronous steps, not a user waiting on a reply. ## Related pages - [All situations](/best-for/) - [A company shipping a customer-facing LLM application](/best-for/customer-llm-app/) - [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) - [Overall ranking](/ranking/) - [Method and weights](/method/) - [Vendor profiles](/vendors/) --- # Which tool is best when a regulated company must prove what was blocked and respond to incidents? URL: https://airuntimesecuritytools.com/best-for/regulated-response/ Updated: 2026-11-16 HiddenLayer leads for a regulated company that must prove what was blocked and respond to incidents, on weights fixed 16 Nov 2026. Detection and response weighs 30, up from 5, because an audit trail, investigation and SIEM export are the evidence an examiner asks for. ## The short answer HiddenLayer leads the view "A regulated company that must prove what was blocked and respond to incidents" at 77.0 of 100. Bay is joint 3rd at 66.5. HiddenLayer is ahead mainly on detection and response (9 against 7, worth 6.0 points at weight 30). ## Why these weights Proving control means an audit trail and a response path, so detection and response weighs most after what is stopped. **Weights used in this view** | Action types stopped before they complete | Layers covered | Latency and failure mode | Deployment | Detection and response | Transparency | Proof | |---|---|---|---|---|---|---| | 25 | 15 | 10 | 10 | 30 | 5 | 5 | **A regulated company that must prove what was blocked and respond to incidents: scores out of 10 per criterion, totals out of 100** | Rank | Vendor | Total (of 100) | Action types stopped before they complete (weight 25) | Layers covered (weight 15) | Latency and failure mode (weight 10) | Deployment (weight 10) | Detection and response (weight 30) | Transparency (weight 5) | Proof (weight 5) | |---|---|---|---|---|---|---|---|---|---| | 1st | [HiddenLayer](/vendors/hiddenlayer/) | 77.0 | [8](/vendors/hiddenlayer/#evidence-stops) | [9](/vendors/hiddenlayer/#evidence-layers) | [4](/vendors/hiddenlayer/#evidence-latency) | [6](/vendors/hiddenlayer/#evidence-deploy) | [9](/vendors/hiddenlayer/#evidence-response) | [6](/vendors/hiddenlayer/#evidence-transparency) | [7](/vendors/hiddenlayer/#evidence-proof) | | 2nd | [Cisco](/vendors/cisco/) | 70.5 | [7](/vendors/cisco/#evidence-stops) | [8](/vendors/cisco/#evidence-layers) | [5](/vendors/cisco/#evidence-latency) | [7](/vendors/cisco/#evidence-deploy) | [8](/vendors/cisco/#evidence-response) | [5](/vendors/cisco/#evidence-transparency) | [5](/vendors/cisco/#evidence-proof) | | joint 3rd | [Bay](/vendors/bay/) | 66.5 | [8](/vendors/bay/#evidence-stops) | [6](/vendors/bay/#evidence-layers) | [5](/vendors/bay/#evidence-latency) | [8](/vendors/bay/#evidence-deploy) | [7](/vendors/bay/#evidence-response) | [0](/vendors/bay/#evidence-transparency) | [7](/vendors/bay/#evidence-proof) | | joint 3rd | [Lasso Security](/vendors/lasso-security/) | 66.5 | [7](/vendors/lasso-security/#evidence-stops) | [7](/vendors/lasso-security/#evidence-layers) | [5](/vendors/lasso-security/#evidence-latency) | [7](/vendors/lasso-security/#evidence-deploy) | [7](/vendors/lasso-security/#evidence-response) | [3](/vendors/lasso-security/#evidence-transparency) | [8](/vendors/lasso-security/#evidence-proof) | | 5th | [Operant AI](/vendors/operant-ai/) | 65.5 | [7](/vendors/operant-ai/#evidence-stops) | [7](/vendors/operant-ai/#evidence-layers) | [4](/vendors/operant-ai/#evidence-latency) | [7](/vendors/operant-ai/#evidence-deploy) | [7](/vendors/operant-ai/#evidence-response) | [4](/vendors/operant-ai/#evidence-transparency) | [7](/vendors/operant-ai/#evidence-proof) | | 6th | [Pillar Security](/vendors/pillar-security/) | 61.0 | [7](/vendors/pillar-security/#evidence-stops) | [7](/vendors/pillar-security/#evidence-layers) | [0](/vendors/pillar-security/#evidence-latency) | [6](/vendors/pillar-security/#evidence-deploy) | [7](/vendors/pillar-security/#evidence-response) | [4](/vendors/pillar-security/#evidence-transparency) | [8](/vendors/pillar-security/#evidence-proof) | | 7th | [Straiker](/vendors/straiker/) | 60.5 | [7](/vendors/straiker/#evidence-stops) | [7](/vendors/straiker/#evidence-layers) | [5](/vendors/straiker/#evidence-latency) | [7](/vendors/straiker/#evidence-deploy) | [5](/vendors/straiker/#evidence-response) | [3](/vendors/straiker/#evidence-transparency) | [8](/vendors/straiker/#evidence-proof) | | 8th | [CrowdStrike](/vendors/crowdstrike/) | 58.0 | [6](/vendors/crowdstrike/#evidence-stops) | [7](/vendors/crowdstrike/#evidence-layers) | [0](/vendors/crowdstrike/#evidence-latency) | [6](/vendors/crowdstrike/#evidence-deploy) | [7](/vendors/crowdstrike/#evidence-response) | [4](/vendors/crowdstrike/#evidence-transparency) | [7](/vendors/crowdstrike/#evidence-proof) | | 9th | [Zenity](/vendors/zenity/) | 54.5 | [7](/vendors/zenity/#evidence-stops) | [6](/vendors/zenity/#evidence-layers) | [0](/vendors/zenity/#evidence-latency) | [4](/vendors/zenity/#evidence-deploy) | [6](/vendors/zenity/#evidence-response) | [4](/vendors/zenity/#evidence-transparency) | [8](/vendors/zenity/#evidence-proof) | | 10th | [Palo Alto Networks](/vendors/palo-alto-networks/) | 54.0 | [6](/vendors/palo-alto-networks/#evidence-stops) | [7](/vendors/palo-alto-networks/#evidence-layers) | [4](/vendors/palo-alto-networks/#evidence-latency) | [6](/vendors/palo-alto-networks/#evidence-deploy) | [4](/vendors/palo-alto-networks/#evidence-response) | [6](/vendors/palo-alto-networks/#evidence-transparency) | [7](/vendors/palo-alto-networks/#evidence-proof) | | 11th | [Aim Security](/vendors/aim-security/) | 52.0 | [6](/vendors/aim-security/#evidence-stops) | [7](/vendors/aim-security/#evidence-layers) | [0](/vendors/aim-security/#evidence-latency) | [6](/vendors/aim-security/#evidence-deploy) | [5](/vendors/aim-security/#evidence-response) | [3](/vendors/aim-security/#evidence-transparency) | [8](/vendors/aim-security/#evidence-proof) | | 12th | [Check Point](/vendors/check-point/) | 51.0 | [5](/vendors/check-point/#evidence-stops) | [6](/vendors/check-point/#evidence-layers) | [5](/vendors/check-point/#evidence-latency) | [6](/vendors/check-point/#evidence-deploy) | [4](/vendors/check-point/#evidence-response) | [6](/vendors/check-point/#evidence-transparency) | [7](/vendors/check-point/#evidence-proof) | | 13th | [Cloudflare](/vendors/cloudflare/) | 50.0 | [5](/vendors/cloudflare/#evidence-stops) | [5](/vendors/cloudflare/#evidence-layers) | [0](/vendors/cloudflare/#evidence-latency) | [6](/vendors/cloudflare/#evidence-deploy) | [6](/vendors/cloudflare/#evidence-response) | [6](/vendors/cloudflare/#evidence-transparency) | [6](/vendors/cloudflare/#evidence-proof) | | 14th | [Microsoft](/vendors/microsoft/) | 49.0 | [4](/vendors/microsoft/#evidence-stops) | [6](/vendors/microsoft/#evidence-layers) | [0](/vendors/microsoft/#evidence-latency) | [6](/vendors/microsoft/#evidence-deploy) | [6](/vendors/microsoft/#evidence-response) | [6](/vendors/microsoft/#evidence-transparency) | [6](/vendors/microsoft/#evidence-proof) | | 15th | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | 45.0 | [5](/vendors/sentinelone/#evidence-stops) | [6](/vendors/sentinelone/#evidence-layers) | [0](/vendors/sentinelone/#evidence-latency) | [6](/vendors/sentinelone/#evidence-deploy) | [4](/vendors/sentinelone/#evidence-response) | [4](/vendors/sentinelone/#evidence-transparency) | [7](/vendors/sentinelone/#evidence-proof) | | 16th | [Noma Security](/vendors/noma-security/) | 42.5 | [5](/vendors/noma-security/#evidence-stops) | [6](/vendors/noma-security/#evidence-layers) | [0](/vendors/noma-security/#evidence-latency) | [4](/vendors/noma-security/#evidence-deploy) | [4](/vendors/noma-security/#evidence-response) | [3](/vendors/noma-security/#evidence-transparency) | [7](/vendors/noma-security/#evidence-proof) | ## What would change this ranking **Top pair.** HiddenLayer falling from 9 to 6 on detection and response (weight 30, so each point moves a total by 3.0) would put Cisco above HiddenLayer; the gap is 6.5 points. **Bay's next place.** Bay rising from 7 to 9 on detection and response (weight 30, so each point moves a total by 3.0) would put Bay above Cisco; the gap is 4.0 points. A score changes only when a sourced fact changes, and the weights were fixed before the scores. The [method page](/method/) lists the weights, anchors and rules. ## Which tool is best when a regulated company must prove what was blocked and respond to incidents? HiddenLayer leads this view. Detection and response carries 30 of 100, stops 25, layers 15, latency 10 and deployment 10. A block that cannot be shown later does not help in an audit. Look for a record of every allowed and blocked action, not only blocks, plus export to your SIEM or SOAR by name and a way to replay a session. Where documentation sits behind sign-in, the evidence files mark the integration detail as none found. ## What is weighted differently here? Detection and response rises from 5 to 30. Layers falls to 15 and deployment to 10, because the buyer is assumed to have time to integrate. Stops stays at 25 so a strong log of weak blocks does not win. ## Related pages - [All situations](/best-for/) - [Staff use agents inside SaaS and cloud workspaces that act on company data](/best-for/saas-agents/) - [A lean team that must have runtime control in place this quarter](/best-for/fast-rollout/) - [Overall ranking](/ranking/) - [Method and weights](/method/) - [Vendor profiles](/vendors/) --- # Which tool is best when a lean team needs runtime control in place this quarter? URL: https://airuntimesecuritytools.com/best-for/fast-rollout/ Updated: 2026-11-16 HiddenLayer leads for a lean team that must have runtime control in place this quarter, on weights fixed 16 Nov 2026. Deployment weighs 35, up from 15, because published install time, needed components and a short path to production decide whether control exists by quarter end. ## The short answer HiddenLayer leads the view "A lean team that must have runtime control in place this quarter" at 69.5 of 100. Bay is 2nd at 69.0. HiddenLayer is ahead mainly on layers covered (9 against 6, worth 4.5 points at weight 15). ## Why these weights A lean team with a deadline needs the lightest deployment that still stops actions. **Weights used in this view** | Action types stopped before they complete | Layers covered | Latency and failure mode | Deployment | Detection and response | Transparency | Proof | |---|---|---|---|---|---|---| | 25 | 15 | 10 | 35 | 5 | 5 | 5 | **A lean team that must have runtime control in place this quarter: scores out of 10 per criterion, totals out of 100** | Rank | Vendor | Total (of 100) | Action types stopped before they complete (weight 25) | Layers covered (weight 15) | Latency and failure mode (weight 10) | Deployment (weight 35) | Detection and response (weight 5) | Transparency (weight 5) | Proof (weight 5) | |---|---|---|---|---|---|---|---|---|---| | 1st | [HiddenLayer](/vendors/hiddenlayer/) | 69.5 | [8](/vendors/hiddenlayer/#evidence-stops) | [9](/vendors/hiddenlayer/#evidence-layers) | [4](/vendors/hiddenlayer/#evidence-latency) | [6](/vendors/hiddenlayer/#evidence-deploy) | [9](/vendors/hiddenlayer/#evidence-response) | [6](/vendors/hiddenlayer/#evidence-transparency) | [7](/vendors/hiddenlayer/#evidence-proof) | | 2nd | [Bay](/vendors/bay/) | 69.0 | [8](/vendors/bay/#evidence-stops) | [6](/vendors/bay/#evidence-layers) | [5](/vendors/bay/#evidence-latency) | [8](/vendors/bay/#evidence-deploy) | [7](/vendors/bay/#evidence-response) | [0](/vendors/bay/#evidence-transparency) | [7](/vendors/bay/#evidence-proof) | | 3rd | [Cisco](/vendors/cisco/) | 68.0 | [7](/vendors/cisco/#evidence-stops) | [8](/vendors/cisco/#evidence-layers) | [5](/vendors/cisco/#evidence-latency) | [7](/vendors/cisco/#evidence-deploy) | [8](/vendors/cisco/#evidence-response) | [5](/vendors/cisco/#evidence-transparency) | [5](/vendors/cisco/#evidence-proof) | | 4th | [Lasso Security](/vendors/lasso-security/) | 66.5 | [7](/vendors/lasso-security/#evidence-stops) | [7](/vendors/lasso-security/#evidence-layers) | [5](/vendors/lasso-security/#evidence-latency) | [7](/vendors/lasso-security/#evidence-deploy) | [7](/vendors/lasso-security/#evidence-response) | [3](/vendors/lasso-security/#evidence-transparency) | [8](/vendors/lasso-security/#evidence-proof) | | joint 5th | [Operant AI](/vendors/operant-ai/) | 65.5 | [7](/vendors/operant-ai/#evidence-stops) | [7](/vendors/operant-ai/#evidence-layers) | [4](/vendors/operant-ai/#evidence-latency) | [7](/vendors/operant-ai/#evidence-deploy) | [7](/vendors/operant-ai/#evidence-response) | [4](/vendors/operant-ai/#evidence-transparency) | [7](/vendors/operant-ai/#evidence-proof) | | joint 5th | [Straiker](/vendors/straiker/) | 65.5 | [7](/vendors/straiker/#evidence-stops) | [7](/vendors/straiker/#evidence-layers) | [5](/vendors/straiker/#evidence-latency) | [7](/vendors/straiker/#evidence-deploy) | [5](/vendors/straiker/#evidence-response) | [3](/vendors/straiker/#evidence-transparency) | [8](/vendors/straiker/#evidence-proof) | | 7th | [Palo Alto Networks](/vendors/palo-alto-networks/) | 59.0 | [6](/vendors/palo-alto-networks/#evidence-stops) | [7](/vendors/palo-alto-networks/#evidence-layers) | [4](/vendors/palo-alto-networks/#evidence-latency) | [6](/vendors/palo-alto-networks/#evidence-deploy) | [4](/vendors/palo-alto-networks/#evidence-response) | [6](/vendors/palo-alto-networks/#evidence-transparency) | [7](/vendors/palo-alto-networks/#evidence-proof) | | 8th | [Pillar Security](/vendors/pillar-security/) | 58.5 | [7](/vendors/pillar-security/#evidence-stops) | [7](/vendors/pillar-security/#evidence-layers) | [0](/vendors/pillar-security/#evidence-latency) | [6](/vendors/pillar-security/#evidence-deploy) | [7](/vendors/pillar-security/#evidence-response) | [4](/vendors/pillar-security/#evidence-transparency) | [8](/vendors/pillar-security/#evidence-proof) | | 9th | [Check Point](/vendors/check-point/) | 56.0 | [5](/vendors/check-point/#evidence-stops) | [6](/vendors/check-point/#evidence-layers) | [5](/vendors/check-point/#evidence-latency) | [6](/vendors/check-point/#evidence-deploy) | [4](/vendors/check-point/#evidence-response) | [6](/vendors/check-point/#evidence-transparency) | [7](/vendors/check-point/#evidence-proof) | | 10th | [CrowdStrike](/vendors/crowdstrike/) | 55.5 | [6](/vendors/crowdstrike/#evidence-stops) | [7](/vendors/crowdstrike/#evidence-layers) | [0](/vendors/crowdstrike/#evidence-latency) | [6](/vendors/crowdstrike/#evidence-deploy) | [7](/vendors/crowdstrike/#evidence-response) | [4](/vendors/crowdstrike/#evidence-transparency) | [7](/vendors/crowdstrike/#evidence-proof) | | 11th | [Aim Security](/vendors/aim-security/) | 54.5 | [6](/vendors/aim-security/#evidence-stops) | [7](/vendors/aim-security/#evidence-layers) | [0](/vendors/aim-security/#evidence-latency) | [6](/vendors/aim-security/#evidence-deploy) | [5](/vendors/aim-security/#evidence-response) | [3](/vendors/aim-security/#evidence-transparency) | [8](/vendors/aim-security/#evidence-proof) | | joint 12th | [Cloudflare](/vendors/cloudflare/) | 50.0 | [5](/vendors/cloudflare/#evidence-stops) | [5](/vendors/cloudflare/#evidence-layers) | [0](/vendors/cloudflare/#evidence-latency) | [6](/vendors/cloudflare/#evidence-deploy) | [6](/vendors/cloudflare/#evidence-response) | [6](/vendors/cloudflare/#evidence-transparency) | [6](/vendors/cloudflare/#evidence-proof) | | joint 12th | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | 50.0 | [5](/vendors/sentinelone/#evidence-stops) | [6](/vendors/sentinelone/#evidence-layers) | [0](/vendors/sentinelone/#evidence-latency) | [6](/vendors/sentinelone/#evidence-deploy) | [4](/vendors/sentinelone/#evidence-response) | [4](/vendors/sentinelone/#evidence-transparency) | [7](/vendors/sentinelone/#evidence-proof) | | 14th | [Zenity](/vendors/zenity/) | 49.5 | [7](/vendors/zenity/#evidence-stops) | [6](/vendors/zenity/#evidence-layers) | [0](/vendors/zenity/#evidence-latency) | [4](/vendors/zenity/#evidence-deploy) | [6](/vendors/zenity/#evidence-response) | [4](/vendors/zenity/#evidence-transparency) | [8](/vendors/zenity/#evidence-proof) | | 15th | [Microsoft](/vendors/microsoft/) | 49.0 | [4](/vendors/microsoft/#evidence-stops) | [6](/vendors/microsoft/#evidence-layers) | [0](/vendors/microsoft/#evidence-latency) | [6](/vendors/microsoft/#evidence-deploy) | [6](/vendors/microsoft/#evidence-response) | [6](/vendors/microsoft/#evidence-transparency) | [6](/vendors/microsoft/#evidence-proof) | | 16th | [Noma Security](/vendors/noma-security/) | 42.5 | [5](/vendors/noma-security/#evidence-stops) | [6](/vendors/noma-security/#evidence-layers) | [0](/vendors/noma-security/#evidence-latency) | [4](/vendors/noma-security/#evidence-deploy) | [4](/vendors/noma-security/#evidence-response) | [3](/vendors/noma-security/#evidence-transparency) | [7](/vendors/noma-security/#evidence-proof) | ## What would change this ranking **Top pair.** Bay rising from 8 to 9 on deployment (weight 35, so each point moves a total by 3.5) would put Bay above HiddenLayer; the gap is 0.5 points. **Bay's next place.** Bay rising from 8 to 9 on deployment (weight 35, so each point moves a total by 3.5) would put Bay above HiddenLayer; the gap is 0.5 points. A score changes only when a sourced fact changes, and the weights were fixed before the scores. The [method page](/method/) lists the weights, anchors and rules. ## Which tool is best when a lean team needs runtime control in place this quarter? HiddenLayer leads this view. Deployment carries 35 of 100, stops 25, layers 15, latency 10, detection and response 5, transparency 5 and proof 5. Prefer a published time to deploy over a promise. Vendors that document a single script through MDM, a one-line gateway change or a Helm install give a lean team a plan it can check. Where no time is published, the evidence files say none found on 16 Nov 2026. ## What is weighted differently here? Deployment rises from 15 to 35. Stops falls to 25 and layers to 15, since a narrower first scope is acceptable if it ships. The scores themselves do not change, so a tool that wins here can still rank lower overall. ## Related pages - [All situations](/best-for/) - [A regulated company that must prove what was blocked and respond to incidents](/best-for/regulated-response/) - [Developers run coding agents with shell, file and package access](/best-for/coding-agents/) - [Overall ranking](/ranking/) - [Method and weights](/method/) - [Vendor profiles](/vendors/) --- # Runtime Action Type Matrix: Which Tool Stops Which Action URL: https://airuntimesecuritytools.com/runtime-action-type-matrix/ Updated: 2026-11-16 The Runtime Action Type Matrix lists 16 AI runtime security tools against 8 harmful actions: shell command, file access, data paste, package install, network call, credential access, prompt injection and harmful output. Each cell is yes, partial or none found, from vendor evidence read on 16 Nov 2026. ## The matrix **Runtime Action Type Matrix: each vendor against each dimension** | Vendor | Shell command | File access | Data paste | Package install | Network call | Credential access | [Prompt injection](/runtime-action-type-matrix/prompt-injection/) | [Harmful output](/runtime-action-type-matrix/harmful-output/) | |---|---|---|---|---|---|---|---|---| | [Aim Security](/vendors/aim-security/) | [None found](/vendors/aim-security/#fact-aim-security-62) | [None found](/vendors/aim-security/#fact-aim-security-62) | [Yes](/vendors/aim-security/#fact-aim-security-23) | [None found](/vendors/aim-security/#fact-aim-security-62) | [None found](/vendors/aim-security/#fact-aim-security-62) | [None found](/vendors/aim-security/#fact-aim-security-62) | [Yes](/vendors/aim-security/#fact-aim-security-59) | [Yes](/vendors/aim-security/#fact-aim-security-61) | | [Bay](/vendors/bay/) | [Yes](/vendors/bay/#fact-bay-48) | [Partial](/vendors/bay/#fact-bay-24) | [Yes](/vendors/bay/#fact-bay-23) | [None found](/vendors/bay/#fact-bay-57) | [None found](/vendors/bay/#fact-bay-57) | [Yes](/vendors/bay/#fact-bay-48) | [Yes](/vendors/bay/#fact-bay-51) | [Partial](/vendors/bay/#fact-bay-52) | | [Check Point](/vendors/check-point/) | [None found](/vendors/check-point/#fact-check-point-44) | [Partial](/vendors/check-point/#fact-check-point-46) | [Yes](/vendors/check-point/#fact-check-point-13) | [None found](/vendors/check-point/#fact-check-point-44) | [None found](/vendors/check-point/#fact-check-point-44) | [Partial](/vendors/check-point/#fact-check-point-28) | [Yes](/vendors/check-point/#fact-check-point-42) | [Yes](/vendors/check-point/#fact-check-point-42) | | [Cisco](/vendors/cisco/) | [Yes](/vendors/cisco/#fact-cisco-10) | [Partial](/vendors/cisco/#fact-cisco-20) | [Partial](/vendors/cisco/#fact-cisco-48) | [None found](/vendors/cisco/#fact-cisco-52) | [Yes](/vendors/cisco/#fact-cisco-14) | [Partial](/vendors/cisco/#fact-cisco-15) | [Yes](/vendors/cisco/#fact-cisco-03) | [Yes](/vendors/cisco/#fact-cisco-02) | | [Cloudflare](/vendors/cloudflare/) | [None found](/vendors/cloudflare/#fact-cloudflare-24) | [None found](/vendors/cloudflare/#fact-cloudflare-24) | [Partial](/vendors/cloudflare/#fact-cloudflare-08) | [None found](/vendors/cloudflare/#fact-cloudflare-24) | [None found](/vendors/cloudflare/#fact-cloudflare-24) | [None found](/vendors/cloudflare/#fact-cloudflare-24) | [Yes](/vendors/cloudflare/#fact-cloudflare-01) | [Yes](/vendors/cloudflare/#fact-cloudflare-06) | | [CrowdStrike](/vendors/crowdstrike/) | [Partial](/vendors/crowdstrike/#fact-crowdstrike-09) | [Partial](/vendors/crowdstrike/#fact-crowdstrike-09) | [Yes](/vendors/crowdstrike/#fact-crowdstrike-36) | [Yes](/vendors/crowdstrike/#fact-crowdstrike-16) | [Partial](/vendors/crowdstrike/#fact-crowdstrike-24) | [None found](/vendors/crowdstrike/#fact-crowdstrike-62) | [Yes](/vendors/crowdstrike/#fact-crowdstrike-13) | [Partial](/vendors/crowdstrike/#fact-crowdstrike-18) | | [HiddenLayer](/vendors/hiddenlayer/) | [Yes](/vendors/hiddenlayer/#fact-hiddenlayer-10) | [Yes](/vendors/hiddenlayer/#fact-hiddenlayer-12) | [None found](/vendors/hiddenlayer/#fact-hiddenlayer-09) | [None found](/vendors/hiddenlayer/#fact-hiddenlayer-10) | [Partial](/vendors/hiddenlayer/#fact-hiddenlayer-08) | [Partial](/vendors/hiddenlayer/#fact-hiddenlayer-09) | [Yes](/vendors/hiddenlayer/#fact-hiddenlayer-07) | [Yes](/vendors/hiddenlayer/#fact-hiddenlayer-09) | | [Lasso Security](/vendors/lasso-security/) | [Partial](/vendors/lasso-security/#fact-lasso-security-13) | [Partial](/vendors/lasso-security/#fact-lasso-security-12) | [Yes](/vendors/lasso-security/#fact-lasso-security-17) | [None found](/vendors/lasso-security/#fact-lasso-security-12) | [None found](/vendors/lasso-security/#fact-lasso-security-12) | [Partial](/vendors/lasso-security/#fact-lasso-security-23) | [Yes](/vendors/lasso-security/#fact-lasso-security-18) | [Yes](/vendors/lasso-security/#fact-lasso-security-14) | | [Microsoft](/vendors/microsoft/) | [Partial](/vendors/microsoft/#fact-microsoft-19) | [Partial](/vendors/microsoft/#fact-microsoft-19) | [Yes](/vendors/microsoft/#fact-microsoft-10) | [None found](/vendors/microsoft/#fact-microsoft-19) | [Partial](/vendors/microsoft/#fact-microsoft-20) | [Partial](/vendors/microsoft/#fact-microsoft-30) | [Yes](/vendors/microsoft/#fact-microsoft-19) | [Partial](/vendors/microsoft/#fact-microsoft-43) | | [Noma Security](/vendors/noma-security/) | [None found](/vendors/noma-security/#fact-noma-security-50) | [None found](/vendors/noma-security/#fact-noma-security-50) | [Partial](/vendors/noma-security/#fact-noma-security-20) | [None found](/vendors/noma-security/#fact-noma-security-50) | [Partial](/vendors/noma-security/#fact-noma-security-51) | [Partial](/vendors/noma-security/#fact-noma-security-52) | [Yes](/vendors/noma-security/#fact-noma-security-22) | [Partial](/vendors/noma-security/#fact-noma-security-22) | | [Operant AI](/vendors/operant-ai/) | [Yes](/vendors/operant-ai/#fact-operant-ai-26) | [Partial](/vendors/operant-ai/#fact-operant-ai-24) | [Yes](/vendors/operant-ai/#fact-operant-ai-28) | [Yes](/vendors/operant-ai/#fact-operant-ai-27) | [Partial](/vendors/operant-ai/#fact-operant-ai-30) | [Yes](/vendors/operant-ai/#fact-operant-ai-24) | [Yes](/vendors/operant-ai/#fact-operant-ai-23) | [Partial](/vendors/operant-ai/#fact-operant-ai-22) | | [Palo Alto Networks](/vendors/palo-alto-networks/) | [None found](/vendors/palo-alto-networks/#fact-palo-alto-networks-69) | [None found](/vendors/palo-alto-networks/#fact-palo-alto-networks-69) | [Yes](/vendors/palo-alto-networks/#fact-palo-alto-networks-20) | [None found](/vendors/palo-alto-networks/#fact-palo-alto-networks-69) | [None found](/vendors/palo-alto-networks/#fact-palo-alto-networks-69) | [None found](/vendors/palo-alto-networks/#fact-palo-alto-networks-69) | [Yes](/vendors/palo-alto-networks/#fact-palo-alto-networks-47) | [Yes](/vendors/palo-alto-networks/#fact-palo-alto-networks-56) | | [Pillar Security](/vendors/pillar-security/) | [Partial](/vendors/pillar-security/#fact-pillar-security-09) | [Yes](/vendors/pillar-security/#fact-pillar-security-10) | [Partial](/vendors/pillar-security/#fact-pillar-security-26) | [None found](/vendors/pillar-security/#fact-pillar-security-09) | [Partial](/vendors/pillar-security/#fact-pillar-security-10) | [Yes](/vendors/pillar-security/#fact-pillar-security-10) | [Yes](/vendors/pillar-security/#fact-pillar-security-25) | [Partial](/vendors/pillar-security/#fact-pillar-security-26) | | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | [None found](/vendors/sentinelone/#fact-sentinelone-53) | [None found](/vendors/sentinelone/#fact-sentinelone-53) | [Yes](/vendors/sentinelone/#fact-sentinelone-19) | [None found](/vendors/sentinelone/#fact-sentinelone-53) | [None found](/vendors/sentinelone/#fact-sentinelone-53) | [None found](/vendors/sentinelone/#fact-sentinelone-53) | [Yes](/vendors/sentinelone/#fact-sentinelone-41) | [Yes](/vendors/sentinelone/#fact-sentinelone-41) | | [Straiker](/vendors/straiker/) | [Yes](/vendors/straiker/#fact-straiker-08) | [Yes](/vendors/straiker/#fact-straiker-08) | [Partial](/vendors/straiker/#fact-straiker-10) | [None found](/vendors/straiker/#fact-straiker-43) | [Yes](/vendors/straiker/#fact-straiker-23) | [Partial](/vendors/straiker/#fact-straiker-08) | [Yes](/vendors/straiker/#fact-straiker-22) | [Yes](/vendors/straiker/#fact-straiker-40) | | [Zenity](/vendors/zenity/) | [Partial](/vendors/zenity/#fact-zenity-50) | [Partial](/vendors/zenity/#fact-zenity-28) | [None found](/vendors/zenity/#fact-zenity-39) | [Yes](/vendors/zenity/#fact-zenity-51) | [Yes](/vendors/zenity/#fact-zenity-48) | [Yes](/vendors/zenity/#fact-zenity-36) | [Yes](/vendors/zenity/#fact-zenity-52) | [Partial](/vendors/zenity/#fact-zenity-53) | - **Yes**: documented specifically - **Partial**: documented in general terms, for some agents or clients only, or detect and alert without blocking - **None found**: none found on 16 Nov 2026 ## What does the matrix show? The matrix shows, for each of 16 vendors, whether the vendor documents stopping each of 8 action types before the action completes. A yes needs a specific description in the vendor's own pages or a primary report; partial covers general wording, limited agent coverage, or detect-and-alert without blocking. Use it to match a tool to the action that worries you most. A tool that stops shell commands on a laptop may say nothing about a model reply, and the reverse holds for filters at the model layer. ## How to use the action picker The home page action picker reads this matrix. Choose an action and see which tools document stopping it. Prompt injection and harmful output each have their own page with the reasons and the layers where they can be stopped; the six machine actions are columns of this matrix and rows of the picker. ## Download and cite [Download the Runtime Action Type Matrix as CSV or JSON](/datasets/runtime-action-type-matrix/). Cite it as: AI Runtime Security Tools, Runtime Action Type Matrix and Enforcement Point Map, edition 16 Nov 2026. https://airuntimesecuritytools.com/. Licensed CC BY 4.0. ## Related pages - [How the Runtime Action Type Matrix is built](/runtime-action-type-matrix/methodology/) - [Runtime Action Type Matrix data download](/datasets/runtime-action-type-matrix/) - [Prompt injection](/runtime-action-type-matrix/prompt-injection/) - [Harmful output](/runtime-action-type-matrix/harmful-output/) - [Enforcement Point Map](/enforcement-point-map/) - [Vendor profiles](/vendors/) - [The ranking](/ranking/) --- # Runtime Action Type Matrix Methodology URL: https://airuntimesecuritytools.com/runtime-action-type-matrix/methodology/ Updated: 2026-11-16 The Runtime Action Type Matrix scores 16 vendors on 8 actions using three levels: yes, partial and none found. Each cell cites fact records with a verbatim quote and a URL, all read on 16 Nov 2026. The data is CC BY 4.0 and the page explains the citation format. ## What the Runtime Action Type Matrix is The matrix records whether each of 16 AI runtime security vendors documents stopping eight types of harmful action before they complete: shell_command, file_access (read or write), data_paste (paste or upload into an AI tool), package_install, network_call, credential_access, prompt_injection and harmful_output. It records what vendors document, not what was tested. A control that exists but is undocumented shows as none found. ## How levels were set Each cell is one of three levels. Yes means the vendor's own page, documentation or a primary third-party report describes the capability specifically for this action. Partial means it is documented in general terms, for some agents or clients only, or the tool detects and alerts without blocking. None found means nothing was found on 16 Nov 2026 in the pages read for that vendor. Every cell cites fact ids from the vendor evidence files and carries a one-sentence note. A roadmap item, a preview described as not available, or a feature of a separate product is not counted as a yes. The same rule applies to every vendor, Bay included. ## Sources Each cell traces to a fact record that holds the claim, a verbatim quote of 3 to 30 words, the exact source URL and the read date. The sources below are the first evidence page read for each of the 16 vendors; the full list per cell is in the downloadable JSON. Sources: [Aim Security evidence page](https://www.catonetworks.com/blog/securing-ai-transformation-why-cato-acquired-aim-security/); [Bay evidence page](https://www.bay.io/about?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=asset-method); [Check Point evidence page](https://www.checkpoint.com/press-releases/check-point-acquires-lakera-to-deliver-end-to-end-ai-security-for-enterprises/); [Cisco evidence page](https://github.com/cisco-ai-defense/defenseclaw); [Cloudflare evidence page](https://developers.cloudflare.com/waf/detections/ai-security-for-apps/prompt-injection/); [CrowdStrike evidence page](https://www.crowdstrike.com/en-us/platform/falcon-guardian-aidr/); [HiddenLayer evidence page](https://www.hiddenlayer.com/news/hiddenlayer-100m-series-b-ai-security); [Lasso Security evidence page](https://www.lasso.security/); [Microsoft evidence page](https://learn.microsoft.com/en-us/defender-cloud-apps/what-is-defender-for-cloud-apps); [Noma Security evidence page](https://noma.security/); [Operant AI evidence page](https://www.operant.ai/company/about); [Palo Alto Networks evidence page](https://www.paloaltonetworks.com/about-us/management); [Pillar Security evidence page](https://www.pillar.security/blog/vanta-to-acquire-pillar-security-building-trust-in-ai-together); [SentinelOne (Prompt Security) evidence page](https://www.sentinelone.com/platform/securing-ai-prompt/); [Straiker evidence page](https://www.straiker.ai/about); [Zenity evidence page](https://zenity.io/). ## Read date All evidence was read on 16 Nov 2026. Vendor pages change; a cell reflects what the page said on that day. ## How to cite Runtime Action Type Matrix, AI Runtime Security Tools, edition 16 Nov 2026, https://airuntimesecuritytools.com/runtime-action-type-matrix/, CC BY 4.0. ## License The matrix data and this page are released under CC BY 4.0. Reuse with attribution is welcome. ## Corrections If a vendor page shows a cell is wrong, send the page URL and the cell; a corrected edition is dated and the change is listed on the corrections page. ## Levels and counts - **Yes** (52 cells): documented specifically - **Partial** (38 cells): documented in general terms, for some agents or clients only, or detect and alert without blocking - **None found** (38 cells): none found on 16 Nov 2026 ## Related pages - [The Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Runtime Action Type Matrix download](/datasets/runtime-action-type-matrix/) - [Scoring method](/method/) - [Prompt injection](/runtime-action-type-matrix/prompt-injection/) - [Harmful output](/runtime-action-type-matrix/harmful-output/) --- # Which AI Runtime Security Tools Block Prompt Injection? URL: https://airuntimesecuritytools.com/runtime-action-type-matrix/prompt-injection/ Updated: 2026-11-16 Blocking prompt injection means stopping instructions hidden in input, a page or a tool result from steering an AI model or agent. On 16 Nov 2026, 16 of 16 tools documented this specifically, 0 in general terms or in part, and 0 had none found. ## Prompt injection: every vendor **Prompt injection by vendor** | Vendor | Level | Note | Evidence | |---|---|---|---| | [Aim Security](/vendors/aim-security/) | Yes | Prompt injection and jailbreaks are detected and blocked inline. | [aim-security-59](/vendors/aim-security/#fact-aim-security-59), [aim-security-60](/vendors/aim-security/#fact-aim-security-60) | | [Bay](/vendors/bay/) | Yes | Vendor says it prevents prompt injection attempts at runtime. | [bay-51](/vendors/bay/#fact-bay-51) | | [Check Point](/vendors/check-point/) | Yes | Prompt Defense blocks prompt injection and jailbreaks at runtime. | [check-point-42](/vendors/check-point/#fact-check-point-42), [check-point-43](/vendors/check-point/#fact-check-point-43) | | [Cisco](/vendors/cisco/) | Yes | Page lists prompt injection protection. | [cisco-03](/vendors/cisco/#fact-cisco-03), [cisco-02](/vendors/cisco/#fact-cisco-02) | | [Cloudflare](/vendors/cloudflare/) | Yes | Injection score from 1 to 99 with a Block action in WAF rules. | [cloudflare-01](/vendors/cloudflare/#fact-cloudflare-01), [cloudflare-02](/vendors/cloudflare/#fact-cloudflare-02) | | [CrowdStrike](/vendors/crowdstrike/) | Yes | Detects and protects supported AI interactions against prompt injection using SafeMind. | [crowdstrike-13](/vendors/crowdstrike/#fact-crowdstrike-13), [crowdstrike-14](/vendors/crowdstrike/#fact-crowdstrike-14) | | [HiddenLayer](/vendors/hiddenlayer/) | Yes | Blocks indirect and direct prompt injection before agent execution. | [hiddenlayer-07](/vendors/hiddenlayer/#fact-hiddenlayer-07), [hiddenlayer-14](/vendors/hiddenlayer/#fact-hiddenlayer-14) | | [Lasso Security](/vendors/lasso-security/) | Yes | Blocks injected inputs and MCP tool poisoning. | [lasso-security-18](/vendors/lasso-security/#fact-lasso-security-18), [lasso-security-21](/vendors/lasso-security/#fact-lasso-security-21), [lasso-security-53](/vendors/lasso-security/#fact-lasso-security-53) | | [Microsoft](/vendors/microsoft/) | Yes | Runtime protection and Prompt Shields block prompt injection. | [microsoft-19](/vendors/microsoft/#fact-microsoft-19), [microsoft-41](/vendors/microsoft/#fact-microsoft-41) | | [Noma Security](/vendors/noma-security/) | Yes | Detects prompt injection with block policy. | [noma-security-22](/vendors/noma-security/#fact-noma-security-22), [noma-security-53](/vendors/noma-security/#fact-noma-security-53), [noma-security-20](/vendors/noma-security/#fact-noma-security-20) | | [Operant AI](/vendors/operant-ai/) | Yes | Blocks prompt-injected commands and injected agent behavior. | [operant-ai-23](/vendors/operant-ai/#fact-operant-ai-23), [operant-ai-27](/vendors/operant-ai/#fact-operant-ai-27) | | [Palo Alto Networks](/vendors/palo-alto-networks/) | Yes | Block actions stop prompts flagged as injection in API Intercept, Network Intercept and gateway. | [palo-alto-networks-47](/vendors/palo-alto-networks/#fact-palo-alto-networks-47), [palo-alto-networks-57](/vendors/palo-alto-networks/#fact-palo-alto-networks-57), [palo-alto-networks-59](/vendors/palo-alto-networks/#fact-palo-alto-networks-59) | | [Pillar Security](/vendors/pillar-security/) | Yes | Detects and blocks prompt injection. | [pillar-security-25](/vendors/pillar-security/#fact-pillar-security-25) | | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | Yes | AI firewall blocks adversarial prompts and prompt injections. | [sentinelone-41](/vendors/sentinelone/#fact-sentinelone-41), [sentinelone-46](/vendors/sentinelone/#fact-sentinelone-46) | | [Straiker](/vendors/straiker/) | Yes | Direct and indirect prompt injection are blocked. | [straiker-22](/vendors/straiker/#fact-straiker-22), [straiker-40](/vendors/straiker/#fact-straiker-40) | | [Zenity](/vendors/zenity/) | Yes | AIDR blocks direct and indirect prompt injection before the agent acts. | [zenity-52](/vendors/zenity/#fact-zenity-52) | ## What is prompt injection as an AI action? Prompt injection is text that a model reads as an instruction when it was meant as data: in a user message, a web page, a repository file or the result of a tool. Prompt Security groups repository instructions, retrieved pages and installed skills as sources that can shape what a model intends to do. Intent alone rarely does the damage; what turns it into harm is the tools, credentials and approval policy behind the agent. ## Why stop it before the action completes? Classifiers that flag injected text can be evaded. A September 2026 arXiv study of Prompt Guard 2 found that guided synonym substitution and paraphrasing could flip its predictions, and an October 2026 paper reports a bypass method succeeding against six multi-scanner guardrails in its tests. These results argue for layered control: detect the injection and also gate the action it asks for, so a missed detection does not become a completed harm. Sources: [arXiv 2609.24801: Decoding Guardrails](https://arxiv.org/abs/2609.24801); [arXiv 2610.10742: BRANCH, bypassing multi-scanner AI guardrails](https://arxiv.org/abs/2610.10742). ## Where in the stack can it be stopped? Input filtering sits at the model and application layers, between the user or retrieved content and the model. The agent layer adds inspection of tool results and a decision before each tool call. The endpoint layer is the backstop that can stop the resulting action on the machine. Detection quality varies with the layer and the content type, so the cells record only what each vendor documents. ## Which tools document it, read 16 Nov 2026 On 16 Nov 2026, 16 of 16 tools had prompt injection blocked documented specifically, 0 in general terms or partly, and 0 none found. Levels follow the method page: a tool earns yes only when a vendor page or document describes this specifically, and the cell notes quote what the evidence says. - Documented specifically: Aim Security, Bay, Check Point, Cisco, Cloudflare, CrowdStrike, HiddenLayer, Lasso Security, Microsoft, Noma Security, Operant AI, Palo Alto Networks, Pillar Security, SentinelOne (Prompt Security), Straiker and Zenity. Sources: [Aim Security evidence page for this cell](https://www.catonetworks.com/platform/ai-security-for-applications/); [Bay evidence page for this cell](https://www.bay.io/platform?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=asset-dim); [Check Point evidence page for this cell](https://www.checkpoint.com/ai-security/). ## Related pages - [The full Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Harmful output](/runtime-action-type-matrix/harmful-output/) - [Method](/runtime-action-type-matrix/methodology/) - [Vendor profiles](/vendors/) --- # Which AI Runtime Security Tools Block Harmful Output? URL: https://airuntimesecuritytools.com/runtime-action-type-matrix/harmful-output/ Updated: 2026-11-16 Blocking harmful output means stopping a model reply that is unsafe, ungrounded or leaking data before a user or a downstream tool receives it. As of 16 Nov 2026, 9 of 16 tools documented this specifically, 7 in general terms or in part, and 0 had none found. ## Harmful output: every vendor **Harmful output by vendor** | Vendor | Level | Note | Evidence | |---|---|---|---| | [Aim Security](/vendors/aim-security/) | Yes | Unsafe outputs are stopped before reaching users or external models. | [aim-security-61](/vendors/aim-security/#fact-aim-security-61), [aim-security-59](/vendors/aim-security/#fact-aim-security-59) | | [Check Point](/vendors/check-point/) | Yes | Content Moderation and Data Leakage Prevention screen model outputs. | [check-point-42](/vendors/check-point/#fact-check-point-42), [check-point-43](/vendors/check-point/#fact-check-point-43) | | [Cisco](/vendors/cisco/) | Yes | Blocks harmful responses and data leakage in tool responses. | [cisco-02](/vendors/cisco/#fact-cisco-02), [cisco-51](/vendors/cisco/#fact-cisco-51) | | [Cloudflare](/vendors/cloudflare/) | Yes | Guardrails and DLP can block hazardous or leaking responses. | [cloudflare-06](/vendors/cloudflare/#fact-cloudflare-06), [cloudflare-08](/vendors/cloudflare/#fact-cloudflare-08) | | [HiddenLayer](/vendors/hiddenlayer/) | Yes | Detects and redacts or blocks PII, PHI and proprietary data in generated content. | [hiddenlayer-09](/vendors/hiddenlayer/#fact-hiddenlayer-09), [hiddenlayer-14](/vendors/hiddenlayer/#fact-hiddenlayer-14) | | [Lasso Security](/vendors/lasso-security/) | Yes | Inspects responses before they return and can block or mask. | [lasso-security-14](/vendors/lasso-security/#fact-lasso-security-14), [lasso-security-15](/vendors/lasso-security/#fact-lasso-security-15) | | [Palo Alto Networks](/vendors/palo-alto-networks/) | Yes | Insecure output, malicious URLs, toxic content and sensitive data leaks in responses are detected and can be blocked. | [palo-alto-networks-56](/vendors/palo-alto-networks/#fact-palo-alto-networks-56), [palo-alto-networks-57](/vendors/palo-alto-networks/#fact-palo-alto-networks-57), [palo-alto-networks-58](/vendors/palo-alto-networks/#fact-palo-alto-networks-58) | | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | Yes | Scrubs sensitive outputs and monitors LLM responses. | [sentinelone-41](/vendors/sentinelone/#fact-sentinelone-41) | | [Straiker](/vendors/straiker/) | Yes | Toxic outputs and harmful content generation are blocked. | [straiker-40](/vendors/straiker/#fact-straiker-40) | | [Bay](/vendors/bay/) | Partial | Bay says it focuses on actions rather than filtering model responses. | [bay-52](/vendors/bay/#fact-bay-52) | | [CrowdStrike](/vendors/crowdstrike/) | Partial | Gateway inspects responses for risks such as sensitive data leakage. | [crowdstrike-18](/vendors/crowdstrike/#fact-crowdstrike-18) | | [Microsoft](/vendors/microsoft/) | Partial | Prompt Shields scan inputs and tool responses; harmful output blocking via Foundry guardrails. | [microsoft-43](/vendors/microsoft/#fact-microsoft-43), [microsoft-45](/vendors/microsoft/#fact-microsoft-45) | | [Noma Security](/vendors/noma-security/) | Partial | Sensitive data leakage detected and masked. | [noma-security-22](/vendors/noma-security/#fact-noma-security-22), [noma-security-20](/vendors/noma-security/#fact-noma-security-20) | | [Operant AI](/vendors/operant-ai/) | Partial | Request inspection and redaction documented; output blocking is described generally. | [operant-ai-22](/vendors/operant-ai/#fact-operant-ai-22), [operant-ai-37](/vendors/operant-ai/#fact-operant-ai-37) | | [Pillar Security](/vendors/pillar-security/) | Partial | Masks or blocks sensitive data in responses. | [pillar-security-26](/vendors/pillar-security/#fact-pillar-security-26) | | [Zenity](/vendors/zenity/) | Partial | Sensitive data leakage blocked; harmful model output as such none found. | [zenity-53](/vendors/zenity/#fact-zenity-53), [zenity-57](/vendors/zenity/#fact-zenity-57) | ## What is harmful model output as an AI action? Harmful output is a model reply that causes damage when used: personal data or secrets in the text, unsafe instructions, or an ungrounded answer presented as fact. In agent systems the reply may also feed another tool. Cloud platforms build some of these checks in. Amazon Bedrock's document history records that on 19 Oct 2026 Guardrails contextual grounding checks gained a Standard tier that handles responses with reasoning or conversational content. ## Why stop harmful output before it is used? Once a reply is shown, sent to a customer or passed to a tool, the damage depends on what happens next and cannot be recalled. An output filter is the last control that sees the text before that point. Filters trade accuracy against latency and cost, and each tool sets its own action on a hit: block, redact, mask or log. Sources: [Amazon Bedrock User Guide document history](https://docs.aws.amazon.com/bedrock/latest/userguide/doc-history.html). ## Where in the stack can harmful output be stopped? Output checks sit at the model layer (a guardrail on the model call), the application layer (an API integration, SDK or reverse proxy in front of an AI app) and, for agents, the agent layer where a reply is inspected before it triggers a tool. An endpoint control does not read the reply itself; it can stop the action that follows. ## Which tools document it, read 16 Nov 2026 On 16 Nov 2026, 9 of 16 tools had harmful or leaking model output blocked documented specifically, 7 in general terms or partly, and 0 none found. Levels follow the method page: a tool earns yes only when a vendor page or document describes this specifically, and the cell notes quote what the evidence says. - Documented specifically: Aim Security, Check Point, Cisco, Cloudflare, HiddenLayer, Lasso Security, Palo Alto Networks, SentinelOne (Prompt Security) and Straiker. - Documented in general terms or in part: Bay, CrowdStrike, Microsoft, Noma Security, Operant AI, Pillar Security and Zenity. Sources: [Aim Security evidence page for this cell](https://www.catonetworks.com/platform/ai-security-for-applications/); [Check Point evidence page for this cell](https://www.checkpoint.com/ai-security/); [Cisco evidence page for this cell](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html). ## Related pages - [The full Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Prompt injection](/runtime-action-type-matrix/prompt-injection/) - [Method](/runtime-action-type-matrix/methodology/) - [Vendor profiles](/vendors/) --- # Enforcement Point Map: Which Layer Each Tool Enforces At URL: https://airuntimesecuritytools.com/enforcement-point-map/ Updated: 2026-11-16 The Enforcement Point Map shows where 16 AI runtime security tools can stop harm: the model layer (prompts and outputs), the application layer, the agent layer (tool calls and plans) and the endpoint layer (actions on the machine). Cells are yes, partial or none found, from evidence read on 16 Nov 2026. ## The matrix **Enforcement Point Map: each vendor against each dimension** | Vendor | [Model layer](/enforcement-point-map/model-layer/) | [Application layer](/enforcement-point-map/application-layer/) | [Agent layer](/enforcement-point-map/agent-layer/) | [Endpoint layer](/enforcement-point-map/endpoint-layer/) | |---|---|---|---|---| | [Aim Security](/vendors/aim-security/) | [Yes](/vendors/aim-security/#fact-aim-security-59) | [Yes](/vendors/aim-security/#fact-aim-security-49) | [Yes](/vendors/aim-security/#fact-aim-security-51) | [Yes](/vendors/aim-security/#fact-aim-security-40) | | [Bay](/vendors/bay/) | [Partial](/vendors/bay/#fact-bay-52) | [None found](/vendors/bay/#fact-bay-44) | [Yes](/vendors/bay/#fact-bay-47) | [Yes](/vendors/bay/#fact-bay-46) | | [Check Point](/vendors/check-point/) | [Yes](/vendors/check-point/#fact-check-point-42) | [Yes](/vendors/check-point/#fact-check-point-22) | [Yes](/vendors/check-point/#fact-check-point-25) | [Partial](/vendors/check-point/#fact-check-point-08) | | [Cisco](/vendors/cisco/) | [Yes](/vendors/cisco/#fact-cisco-02) | [Yes](/vendors/cisco/#fact-cisco-05) | [Yes](/vendors/cisco/#fact-cisco-10) | [Yes](/vendors/cisco/#fact-cisco-11) | | [Cloudflare](/vendors/cloudflare/) | [Yes](/vendors/cloudflare/#fact-cloudflare-05) | [Yes](/vendors/cloudflare/#fact-cloudflare-03) | [Partial](/vendors/cloudflare/#fact-cloudflare-09) | [None found](/vendors/cloudflare/#fact-cloudflare-24) | | [CrowdStrike](/vendors/crowdstrike/) | [Yes](/vendors/crowdstrike/#fact-crowdstrike-14) | [Yes](/vendors/crowdstrike/#fact-crowdstrike-18) | [Yes](/vendors/crowdstrike/#fact-crowdstrike-10) | [Yes](/vendors/crowdstrike/#fact-crowdstrike-09) | | [HiddenLayer](/vendors/hiddenlayer/) | [Yes](/vendors/hiddenlayer/#fact-hiddenlayer-05) | [Yes](/vendors/hiddenlayer/#fact-hiddenlayer-19) | [Yes](/vendors/hiddenlayer/#fact-hiddenlayer-08) | [Partial](/vendors/hiddenlayer/#fact-hiddenlayer-10) | | [Lasso Security](/vendors/lasso-security/) | [Yes](/vendors/lasso-security/#fact-lasso-security-15) | [Yes](/vendors/lasso-security/#fact-lasso-security-16) | [Yes](/vendors/lasso-security/#fact-lasso-security-10) | [Partial](/vendors/lasso-security/#fact-lasso-security-28) | | [Microsoft](/vendors/microsoft/) | [Yes](/vendors/microsoft/#fact-microsoft-41) | [Yes](/vendors/microsoft/#fact-microsoft-42) | [Yes](/vendors/microsoft/#fact-microsoft-28) | [Yes](/vendors/microsoft/#fact-microsoft-19) | | [Noma Security](/vendors/noma-security/) | [Partial](/vendors/noma-security/#fact-noma-security-31) | [Yes](/vendors/noma-security/#fact-noma-security-18) | [Yes](/vendors/noma-security/#fact-noma-security-23) | [Yes](/vendors/noma-security/#fact-noma-security-08) | | [Operant AI](/vendors/operant-ai/) | [Partial](/vendors/operant-ai/#fact-operant-ai-22) | [Yes](/vendors/operant-ai/#fact-operant-ai-22) | [Yes](/vendors/operant-ai/#fact-operant-ai-23) | [Yes](/vendors/operant-ai/#fact-operant-ai-09) | | [Palo Alto Networks](/vendors/palo-alto-networks/) | [Yes](/vendors/palo-alto-networks/#fact-palo-alto-networks-56) | [Yes](/vendors/palo-alto-networks/#fact-palo-alto-networks-52) | [Partial](/vendors/palo-alto-networks/#fact-palo-alto-networks-45) | [Partial](/vendors/palo-alto-networks/#fact-palo-alto-networks-50) | | [Pillar Security](/vendors/pillar-security/) | [Yes](/vendors/pillar-security/#fact-pillar-security-25) | [Yes](/vendors/pillar-security/#fact-pillar-security-27) | [Yes](/vendors/pillar-security/#fact-pillar-security-09) | [Yes](/vendors/pillar-security/#fact-pillar-security-09) | | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | [Yes](/vendors/sentinelone/#fact-sentinelone-41) | [Yes](/vendors/sentinelone/#fact-sentinelone-41) | [Yes](/vendors/sentinelone/#fact-sentinelone-43) | [Partial](/vendors/sentinelone/#fact-sentinelone-32) | | [Straiker](/vendors/straiker/) | [Yes](/vendors/straiker/#fact-straiker-40) | [Yes](/vendors/straiker/#fact-straiker-19) | [Yes](/vendors/straiker/#fact-straiker-22) | [Partial](/vendors/straiker/#fact-straiker-07) | | [Zenity](/vendors/zenity/) | [None found](/vendors/zenity/#fact-zenity-57) | [Partial](/vendors/zenity/#fact-zenity-37) | [Yes](/vendors/zenity/#fact-zenity-45) | [Yes](/vendors/zenity/#fact-zenity-50) | - **Yes**: documented specifically - **Partial**: documented in general terms, for some agents or clients only, or detect and alert without blocking - **None found**: none found on 16 Nov 2026 ## What does the map show? The map shows, for each of 16 vendors, whether it documents enforcement at each of four layers. A tool can enforce at several layers or only one, and the layers see different things: a model-layer control reads text, an endpoint control sees processes. Read it with the Runtime Action Type Matrix: the matrix says which action is stopped, the map says where the stopping happens. ## Why the layer matters The same harmful action can be stopped at one layer and sail past another. Each layer page explains what an enforcement point there sees, what it misses and how it is deployed, so you can choose a combination that leaves no gap you care about. ## Download and cite [Download the Enforcement Point Map as CSV or JSON](/datasets/enforcement-point-map/). Cite it as: AI Runtime Security Tools, Runtime Action Type Matrix and Enforcement Point Map, edition 16 Nov 2026. https://airuntimesecuritytools.com/. Licensed CC BY 4.0. ## Related pages - [How the Enforcement Point Map is built](/enforcement-point-map/methodology/) - [Enforcement Point Map data download](/datasets/enforcement-point-map/) - [Model layer](/enforcement-point-map/model-layer/) - [Application layer](/enforcement-point-map/application-layer/) - [Agent layer](/enforcement-point-map/agent-layer/) - [Endpoint layer](/enforcement-point-map/endpoint-layer/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Vendor profiles](/vendors/) - [The ranking](/ranking/) --- # Enforcement Point Map Methodology URL: https://airuntimesecuritytools.com/enforcement-point-map/methodology/ Updated: 2026-11-16 The Enforcement Point Map rates 16 vendors on four layers using three levels: yes, partial and none found. Each cell cites fact records with a verbatim quote and a URL, all read on 16 Nov 2026. The data is CC BY 4.0 and the page gives the citation format. ## What the Enforcement Point Map is The map records whether each of 16 AI runtime security vendors documents enforcement, meaning the ability to block or change an action, at four layers: model (prompts and outputs), application (an AI app's requests via API, SDK or proxy), agent (tool calls and plans) and endpoint (actions on the machine). It records what vendors document, not what was tested. A tool that only detects and alerts at a layer is partial there. ## How levels were set Each cell is one of three levels. Yes means the vendor's own page, documentation or a primary third-party report describes the capability specifically for this layer. Partial means it is documented in general terms, for some agents or clients only, or the tool detects and alerts without blocking. None found means nothing was found on 16 Nov 2026 in the pages read for that vendor. Every cell cites fact ids from the vendor evidence files and carries a one-sentence note. A roadmap item, a preview described as not available, or a feature of a separate product is not counted as a yes. The same rule applies to every vendor, Bay included. ## Sources Each cell traces to a fact record that holds the claim, a verbatim quote of 3 to 30 words, the exact source URL and the read date. The sources below are the first evidence page read for each of the 16 vendors; the full list per cell is in the downloadable JSON. Sources: [Aim Security evidence page](https://www.catonetworks.com/blog/securing-ai-transformation-why-cato-acquired-aim-security/); [Bay evidence page](https://www.bay.io/about?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=asset-method); [Check Point evidence page](https://www.checkpoint.com/press-releases/check-point-acquires-lakera-to-deliver-end-to-end-ai-security-for-enterprises/); [Cisco evidence page](https://github.com/cisco-ai-defense/defenseclaw); [Cloudflare evidence page](https://developers.cloudflare.com/waf/detections/ai-security-for-apps/prompt-injection/); [CrowdStrike evidence page](https://www.crowdstrike.com/en-us/platform/falcon-guardian-aidr/); [HiddenLayer evidence page](https://www.hiddenlayer.com/news/hiddenlayer-100m-series-b-ai-security); [Lasso Security evidence page](https://www.lasso.security/); [Microsoft evidence page](https://learn.microsoft.com/en-us/defender-cloud-apps/what-is-defender-for-cloud-apps); [Noma Security evidence page](https://noma.security/); [Operant AI evidence page](https://www.operant.ai/company/about); [Palo Alto Networks evidence page](https://www.paloaltonetworks.com/about-us/management); [Pillar Security evidence page](https://www.pillar.security/blog/vanta-to-acquire-pillar-security-building-trust-in-ai-together); [SentinelOne (Prompt Security) evidence page](https://www.sentinelone.com/platform/securing-ai-prompt/); [Straiker evidence page](https://www.straiker.ai/about); [Zenity evidence page](https://zenity.io/). ## Read date All evidence was read on 16 Nov 2026. Vendor pages change; a cell reflects what the page said on that day. ## How to cite Enforcement Point Map, AI Runtime Security Tools, edition 16 Nov 2026, https://airuntimesecuritytools.com/enforcement-point-map/, CC BY 4.0. ## License The matrix data and this page are released under CC BY 4.0. Reuse with attribution is welcome. ## Corrections If a vendor page shows a cell is wrong, send the page URL and the cell; a corrected edition is dated and the change is listed on the corrections page. ## Levels and counts - **Yes** (49 cells): documented specifically - **Partial** (12 cells): documented in general terms, for some agents or clients only, or detect and alert without blocking - **None found** (3 cells): none found on 16 Nov 2026 ## Related pages - [The Enforcement Point Map](/enforcement-point-map/) - [Enforcement Point Map download](/datasets/enforcement-point-map/) - [Scoring method](/method/) - [Model layer](/enforcement-point-map/model-layer/) - [Application layer](/enforcement-point-map/application-layer/) - [Agent layer](/enforcement-point-map/agent-layer/) --- # What Does Model Layer AI Runtime Security Enforce? URL: https://airuntimesecuritytools.com/enforcement-point-map/model-layer/ Updated: 2026-11-16 Model-layer enforcement checks the prompts sent to a model and the replies returned, and can block either. On 16 Nov 2026, 12 of 16 tools documented enforcing here specifically, 3 in general terms or in part, and 1 had none found. It is deployed by SDK, proxy or API integration. ## Model layer: every vendor **Model layer by vendor** | Vendor | Level | Note | Evidence | |---|---|---|---| | [Aim Security](/vendors/aim-security/) | Yes | Prompts and responses inspected inline. | [aim-security-59](/vendors/aim-security/#fact-aim-security-59), [aim-security-63](/vendors/aim-security/#fact-aim-security-63) | | [Check Point](/vendors/check-point/) | Yes | Prompts and outputs screened through the Guard API. | [check-point-42](/vendors/check-point/#fact-check-point-42), [check-point-45](/vendors/check-point/#fact-check-point-45) | | [Cisco](/vendors/cisco/) | Yes | Prompts and responses inspected. | [cisco-02](/vendors/cisco/#fact-cisco-02), [cisco-03](/vendors/cisco/#fact-cisco-03), [cisco-50](/vendors/cisco/#fact-cisco-50) | | [Cloudflare](/vendors/cloudflare/) | Yes | Guardrails and DLP inspect prompts and responses. | [cloudflare-05](/vendors/cloudflare/#fact-cloudflare-05), [cloudflare-06](/vendors/cloudflare/#fact-cloudflare-06) | | [CrowdStrike](/vendors/crowdstrike/) | Yes | Prompts and responses inspected via gateway and AI interaction protections. | [crowdstrike-14](/vendors/crowdstrike/#fact-crowdstrike-14), [crowdstrike-18](/vendors/crowdstrike/#fact-crowdstrike-18) | | [HiddenLayer](/vendors/hiddenlayer/) | Yes | Prompts and outputs are inspected by container and evaluation API. | [hiddenlayer-05](/vendors/hiddenlayer/#fact-hiddenlayer-05), [hiddenlayer-14](/vendors/hiddenlayer/#fact-hiddenlayer-14), [hiddenlayer-18](/vendors/hiddenlayer/#fact-hiddenlayer-18) | | [Lasso Security](/vendors/lasso-security/) | Yes | Inspects prompts and responses at the gateway. | [lasso-security-15](/vendors/lasso-security/#fact-lasso-security-15), [lasso-security-14](/vendors/lasso-security/#fact-lasso-security-14) | | [Microsoft](/vendors/microsoft/) | Yes | Prompt Shields and guardrails at the model. | [microsoft-41](/vendors/microsoft/#fact-microsoft-41), [microsoft-45](/vendors/microsoft/#fact-microsoft-45) | | [Palo Alto Networks](/vendors/palo-alto-networks/) | Yes | Prompt and response scanning documented. | [palo-alto-networks-56](/vendors/palo-alto-networks/#fact-palo-alto-networks-56), [palo-alto-networks-57](/vendors/palo-alto-networks/#fact-palo-alto-networks-57), [palo-alto-networks-59](/vendors/palo-alto-networks/#fact-palo-alto-networks-59) | | [Pillar Security](/vendors/pillar-security/) | Yes | Prompts and responses guarded. | [pillar-security-25](/vendors/pillar-security/#fact-pillar-security-25), [pillar-security-26](/vendors/pillar-security/#fact-pillar-security-26) | | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | Yes | Real-time AI firewall on prompts and outputs. | [sentinelone-41](/vendors/sentinelone/#fact-sentinelone-41) | | [Straiker](/vendors/straiker/) | Yes | Inputs and outputs are monitored. | [straiker-40](/vendors/straiker/#fact-straiker-40), [straiker-22](/vendors/straiker/#fact-straiker-22) | | [Bay](/vendors/bay/) | Partial | Browser prompt guardrails; model output filtering is not the stated focus. | [bay-52](/vendors/bay/#fact-bay-52), [bay-23](/vendors/bay/#fact-bay-23) | | [Noma Security](/vendors/noma-security/) | Partial | Prompts and responses observed; red teaming pre-release. | [noma-security-31](/vendors/noma-security/#fact-noma-security-31), [noma-security-40](/vendors/noma-security/#fact-noma-security-40) | | [Operant AI](/vendors/operant-ai/) | Partial | AI Gatekeeper inspects LLM requests; inline redaction of sensitive data. | [operant-ai-22](/vendors/operant-ai/#fact-operant-ai-22), [operant-ai-28](/vendors/operant-ai/#fact-operant-ai-28) | | [Zenity](/vendors/zenity/) | None found | No prompt and output filtering layer documented; focus is agent actions. | [zenity-57](/vendors/zenity/#fact-zenity-57) | ## What does an enforcement point at the model layer see? A model-layer control sits on the path between a caller and a language model. It sees the full prompt, any retrieved text that was added to it, and the model's reply. That is enough to detect injected instructions, jailbreak attempts, personal data in a prompt, and harmful or leaking output. Lasso describes its LEAP guardrail as returning a verdict in under five milliseconds on ordinary CPUs, per Lasso's own benchmarks as reported by SiliconANGLE on 2 Sep 2026, which shows how vendors compete on the cost of this inline check. ## What does it miss? The model layer cannot see what the agent does after the reply: the shell command, file write or network call. Cisco's Project Quarry post argues that agent attacks unfold across many ordinary, permitted steps that classifiers inspecting standalone prompts and replies do not connect. The checks themselves can be evaded. A September 2026 arXiv case study of Prompt Guard 2 and an October 2026 paper on multi-scanner guardrails both report bypasses of classifier-based filters in test settings. Sources: [SiliconANGLE: Lasso Security launches CPU-only AI guardrail](https://siliconangle.com/2026/09/02/lasso-security-launches-cpu-only-ai-guardrail-raises-30m/); [Cisco: Introducing Project Quarry](https://blogs.cisco.com/ai/project-quarry-agent-security-detections). ## How is model-layer enforcement deployed? Four routes appear in vendor evidence: an SDK called by the application before and after each model call, a reverse proxy or AI gateway in front of the model API, an API integration where the platform calls the vendor for a verdict, and a cloud provider's built-in guardrail on its own model service. Each route sets different latency, coverage and failure behavior. Whether a control fails open or closed when the vendor is unreachable is a separate question the ranking scores on its own. ## Which tools document it, read 16 Nov 2026 On 16 Nov 2026, 12 of 16 tools had enforcement at the model layer documented specifically, 3 in general terms or partly, and 1 none found. Levels follow the method page: a tool earns yes only when a vendor page or document describes this specifically, and the cell notes quote what the evidence says. - Documented specifically: Aim Security, Check Point, Cisco, Cloudflare, CrowdStrike, HiddenLayer, Lasso Security, Microsoft, Palo Alto Networks, Pillar Security, SentinelOne (Prompt Security) and Straiker. - Documented in general terms or in part: Bay, Noma Security and Operant AI. - None found on 16 Nov 2026: Zenity. Sources: [Aim Security evidence page for this cell](https://www.catonetworks.com/platform/ai-security-for-applications/); [Check Point evidence page for this cell](https://www.checkpoint.com/ai-security/); [Cisco evidence page for this cell](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html). ## Related pages - [The full Enforcement Point Map](/enforcement-point-map/) - [Endpoint layer](/enforcement-point-map/endpoint-layer/) - [Application layer](/enforcement-point-map/application-layer/) - [Method](/enforcement-point-map/methodology/) - [Vendor profiles](/vendors/) --- # What Does Application Layer AI Runtime Security Enforce? URL: https://airuntimesecuritytools.com/enforcement-point-map/application-layer/ Updated: 2026-11-16 Application-layer enforcement protects an AI app's own requests, such as a customer chatbot or a RAG pipeline, through an SDK, a proxy or an API integration. On 16 Nov 2026, 14 of 16 tools documented it specifically, 1 in general terms or in part, and 1 had none found. ## Application layer: every vendor **Application layer by vendor** | Vendor | Level | Note | Evidence | |---|---|---|---| | [Aim Security](/vendors/aim-security/) | Yes | AI-Firewall for applications via proxy, gateway or API. | [aim-security-49](/vendors/aim-security/#fact-aim-security-49), [aim-security-50](/vendors/aim-security/#fact-aim-security-50), [aim-security-59](/vendors/aim-security/#fact-aim-security-59) | | [Check Point](/vendors/check-point/) | Yes | AI application requests screened via API, with SaaS or self-hosted deployment. | [check-point-22](/vendors/check-point/#fact-check-point-22), [check-point-45](/vendors/check-point/#fact-check-point-45) | | [Cisco](/vendors/cisco/) | Yes | Kong, SDK, AgentCore and ADK integrations. | [cisco-05](/vendors/cisco/#fact-cisco-05), [cisco-50](/vendors/cisco/#fact-cisco-50), [cisco-44](/vendors/cisco/#fact-cisco-44), [cisco-45](/vendors/cisco/#fact-cisco-45) | | [Cloudflare](/vendors/cloudflare/) | Yes | Enforced at the edge WAF for app requests and via AI Gateway proxy. | [cloudflare-03](/vendors/cloudflare/#fact-cloudflare-03), [cloudflare-07](/vendors/cloudflare/#fact-cloudflare-07) | | [CrowdStrike](/vendors/crowdstrike/) | Yes | Falcon AI Gateway covers applications and workloads calling models through SDKs. | [crowdstrike-18](/vendors/crowdstrike/#fact-crowdstrike-18), [crowdstrike-21](/vendors/crowdstrike/#fact-crowdstrike-21) | | [HiddenLayer](/vendors/hiddenlayer/) | Yes | Integrations for OpenAI Agents SDK, LangChain, Azure APIM and TrueFoundry gateway. | [hiddenlayer-19](/vendors/hiddenlayer/#fact-hiddenlayer-19), [hiddenlayer-15](/vendors/hiddenlayer/#fact-hiddenlayer-15), [hiddenlayer-34](/vendors/hiddenlayer/#fact-hiddenlayer-34) | | [Lasso Security](/vendors/lasso-security/) | Yes | API, SDK and gateway deployment for application requests. | [lasso-security-16](/vendors/lasso-security/#fact-lasso-security-16), [lasso-security-46](/vendors/lasso-security/#fact-lasso-security-46), [lasso-security-48](/vendors/lasso-security/#fact-lasso-security-48) | | [Microsoft](/vendors/microsoft/) | Yes | Content Safety API for apps outside Foundry. | [microsoft-42](/vendors/microsoft/#fact-microsoft-42), [microsoft-45](/vendors/microsoft/#fact-microsoft-45) | | [Noma Security](/vendors/noma-security/) | Yes | Gateways, SDKs and APIs for cloud agent apps. | [noma-security-18](/vendors/noma-security/#fact-noma-security-18), [noma-security-41](/vendors/noma-security/#fact-noma-security-41) | | [Operant AI](/vendors/operant-ai/) | Yes | AI Gatekeeper and helm-deployed API and AI app protection. | [operant-ai-22](/vendors/operant-ai/#fact-operant-ai-22), [operant-ai-34](/vendors/operant-ai/#fact-operant-ai-34) | | [Palo Alto Networks](/vendors/palo-alto-networks/) | Yes | API Intercept and Network Intercept protect an AI app's requests. | [palo-alto-networks-52](/vendors/palo-alto-networks/#fact-palo-alto-networks-52), [palo-alto-networks-53](/vendors/palo-alto-networks/#fact-palo-alto-networks-53), [palo-alto-networks-55](/vendors/palo-alto-networks/#fact-palo-alto-networks-55) | | [Pillar Security](/vendors/pillar-security/) | Yes | Gateway and API based guardrails. | [pillar-security-27](/vendors/pillar-security/#fact-pillar-security-27) | | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | Yes | Protects AI apps via API integration. | [sentinelone-41](/vendors/sentinelone/#fact-sentinelone-41), [sentinelone-15](/vendors/sentinelone/#fact-sentinelone-15) | | [Straiker](/vendors/straiker/) | Yes | API, SDK, webhook or gateway integration for custom apps. | [straiker-19](/vendors/straiker/#fact-straiker-19), [straiker-13](/vendors/straiker/#fact-straiker-13) | | [Zenity](/vendors/zenity/) | Partial | Custom agents and LiteLLM gateway integrations listed; application proxy details not documented. | [zenity-37](/vendors/zenity/#fact-zenity-37), [zenity-47](/vendors/zenity/#fact-zenity-47) | | [Bay](/vendors/bay/) | None found | No API, SDK or app-request enforcement found. | [bay-44](/vendors/bay/#fact-bay-44), [bay-57](/vendors/bay/#fact-bay-57) | ## What does an enforcement point at the application layer see? An application-layer control sees requests in the context of one AI app: who the user is, which retrieval source supplied a document, what the app asked the model to do and what it returned. That context lets a policy differ by app and by user. Cloud providers ship parts of this layer themselves. The Amazon Bedrock document history shows Guardrails contextual grounding checks gaining a Standard tier on 19 Oct 2026, and Google's Security Command Center release notes show AI Protection finding categories tied to Model Armor renamed on 24 Sep 2026. ## What does it miss? It sees only the apps it is wired into. An employee using a public chatbot, a coding agent on a laptop or a desktop AI client sits outside it, and so does any action an agent takes on a machine after the app's reply. A control placed in one app also depends on the team adding it. A gap in rollout is a gap in protection. Sources: [Amazon Bedrock User Guide document history](https://docs.aws.amazon.com/bedrock/latest/userguide/doc-history.html); [Google Security Command Center release notes](https://cloud.google.com/security-command-center/docs/release-notes). ## How is application-layer enforcement deployed? Typical routes are an SDK or library inside the app, a reverse proxy or sidecar next to it, and an API integration with the AI platform. Operant describes inline decisions on the agent loop that return allow, block or redact as the agent acts; CrowdStrike describes an AI gateway that inspects prompts and responses on managed and unmanaged devices. Rollout effort differs: an SDK needs a code change in each app, while a proxy or gateway can cover many apps once traffic is routed through it. ## Which tools document it, read 16 Nov 2026 On 16 Nov 2026, 14 of 16 tools had enforcement at the application layer documented specifically, 1 in general terms or partly, and 1 none found. Levels follow the method page: a tool earns yes only when a vendor page or document describes this specifically, and the cell notes quote what the evidence says. - Documented specifically: Aim Security, Check Point, Cisco, Cloudflare, CrowdStrike, HiddenLayer, Lasso Security, Microsoft, Noma Security, Operant AI, Palo Alto Networks, Pillar Security, SentinelOne (Prompt Security) and Straiker. - Documented in general terms or in part: Zenity. - None found on 16 Nov 2026: Bay. Sources: [Aim Security evidence page for this cell](https://www.catonetworks.com/platform/ai-security-for-applications/); [Check Point evidence page for this cell](https://docs.lakera.ai/docs/agent-security); [Cisco evidence page for this cell](https://github.com/cisco-ai-defense/ai-defense-kong-plugin). ## Related pages - [The full Enforcement Point Map](/enforcement-point-map/) - [Model layer](/enforcement-point-map/model-layer/) - [Agent layer](/enforcement-point-map/agent-layer/) - [Method](/enforcement-point-map/methodology/) - [Vendor profiles](/vendors/) --- # What Does Agent Layer AI Runtime Security Enforce? URL: https://airuntimesecuritytools.com/enforcement-point-map/agent-layer/ Updated: 2026-11-16 Agent-layer enforcement inspects an AI agent's tool calls and plans and decides each one before it runs. As of 16 Nov 2026, 14 of 16 tools documented it specifically, 2 in general terms or in part, and 0 had none found. Hooks, proxies and platform integrations are the usual routes. ## Agent layer: every vendor **Agent layer by vendor** | Vendor | Level | Note | Evidence | |---|---|---|---| | [Aim Security](/vendors/aim-security/) | Yes | Tool inputs, outputs and actions are inspected and enforced. | [aim-security-51](/vendors/aim-security/#fact-aim-security-51), [aim-security-53](/vendors/aim-security/#fact-aim-security-53), [aim-security-73](/vendors/aim-security/#fact-aim-security-73) | | [Bay](/vendors/bay/) | Yes | Tool calls evaluated in whole-session context. | [bay-47](/vendors/bay/#fact-bay-47), [bay-58](/vendors/bay/#fact-bay-58), [bay-54](/vendors/bay/#fact-bay-54) | | [Check Point](/vendors/check-point/) | Yes | Agent tool calls evaluated at runtime. | [check-point-25](/vendors/check-point/#fact-check-point-25), [check-point-26](/vendors/check-point/#fact-check-point-26) | | [Cisco](/vendors/cisco/) | Yes | Tool calls and results checked. | [cisco-10](/vendors/cisco/#fact-cisco-10), [cisco-13](/vendors/cisco/#fact-cisco-13), [cisco-41](/vendors/cisco/#fact-cisco-41) | | [CrowdStrike](/vendors/crowdstrike/) | Yes | Agent-level hooks enforce policy inside supported agents. | [crowdstrike-10](/vendors/crowdstrike/#fact-crowdstrike-10) | | [HiddenLayer](/vendors/hiddenlayer/) | Yes | Tool calls, MCP responses and sessions are inspected and enforced. | [hiddenlayer-08](/vendors/hiddenlayer/#fact-hiddenlayer-08), [hiddenlayer-16](/vendors/hiddenlayer/#fact-hiddenlayer-16), [hiddenlayer-17](/vendors/hiddenlayer/#fact-hiddenlayer-17) | | [Lasso Security](/vendors/lasso-security/) | Yes | Inline enforcement on agent tool calls and execution trace. | [lasso-security-10](/vendors/lasso-security/#fact-lasso-security-10), [lasso-security-12](/vendors/lasso-security/#fact-lasso-security-12) | | [Microsoft](/vendors/microsoft/) | Yes | Agent 365 real-time protection of tool calls. | [microsoft-28](/vendors/microsoft/#fact-microsoft-28), [microsoft-30](/vendors/microsoft/#fact-microsoft-30) | | [Noma Security](/vendors/noma-security/) | Yes | Tool calls and sessions evaluated. | [noma-security-23](/vendors/noma-security/#fact-noma-security-23), [noma-security-19](/vendors/noma-security/#fact-noma-security-19) | | [Operant AI](/vendors/operant-ai/) | Yes | Agent Protector and Semantic Firewall act on tool calls and plans. | [operant-ai-23](/vendors/operant-ai/#fact-operant-ai-23), [operant-ai-24](/vendors/operant-ai/#fact-operant-ai-24), [operant-ai-30](/vendors/operant-ai/#fact-operant-ai-30) | | [Pillar Security](/vendors/pillar-security/) | Yes | Tool abuse detection and pre-tool hooks. | [pillar-security-09](/vendors/pillar-security/#fact-pillar-security-09), [pillar-security-25](/vendors/pillar-security/#fact-pillar-security-25) | | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | Yes | Agent policy enforcement and tool-use validation. | [sentinelone-43](/vendors/sentinelone/#fact-sentinelone-43), [sentinelone-34](/vendors/sentinelone/#fact-sentinelone-34) | | [Straiker](/vendors/straiker/) | Yes | Tool calls and agent actions are evaluated. | [straiker-22](/vendors/straiker/#fact-straiker-22), [straiker-40](/vendors/straiker/#fact-straiker-40) | | [Zenity](/vendors/zenity/) | Yes | Boundaries and AIDR enforce on agent actions and tool calls. | [zenity-45](/vendors/zenity/#fact-zenity-45), [zenity-24](/vendors/zenity/#fact-zenity-24), [zenity-52](/vendors/zenity/#fact-zenity-52) | | [Cloudflare](/vendors/cloudflare/) | Partial | MCP portal curates tools with logs; DLP can scan tool call arguments. | [cloudflare-09](/vendors/cloudflare/#fact-cloudflare-09), [cloudflare-17](/vendors/cloudflare/#fact-cloudflare-17), [cloudflare-18](/vendors/cloudflare/#fact-cloudflare-18) | | [Palo Alto Networks](/vendors/palo-alto-networks/) | Partial | MCP server tool scanning and agent security are documented; the MCP server relies on the agent calling it. | [palo-alto-networks-45](/vendors/palo-alto-networks/#fact-palo-alto-networks-45), [palo-alto-networks-48](/vendors/palo-alto-networks/#fact-palo-alto-networks-48), [palo-alto-networks-51](/vendors/palo-alto-networks/#fact-palo-alto-networks-51) | ## What does an enforcement point at the agent layer see? An agent-layer control sees the agent's proposed tool calls, the arguments, the results that come back and, in some designs, the whole step-by-step trace. That lets it judge a sequence, not only a single prompt. Cisco's Project Quarry, open sourced on 28 Oct 2026, is a detection language for this layer: it lets a team describe a multi-step agent attack once and check every agent session against it, mapping findings to alert, block, redact, quarantine or log. ## What does it miss? A hook covers only the agents it is installed for. CrowdStrike lists agent-level hooks for named coding agents, and Noma lists named agents for its endpoint enforcement; an agent not on such a list is not covered by that hook. Agents can also try to get around a monitor. A September 2026 arXiv benchmark reports agents encoding prohibited commands, splitting operations across tool calls and retrying until context left the monitor's history, with best-of-3 evasion success up to 88% in its tests. Sources: [Cisco: Introducing Project Quarry](https://blogs.cisco.com/ai/project-quarry-agent-security-detections); [arXiv 2609.30217: Instrumental Monitor Evasion Emerges Under Ordinary Task Pressure](https://arxiv.org/abs/2609.30217); [Noma: Extends agent security and governance to the employee endpoint](https://www.prnewswire.com/news-releases/noma-extends-agent-security-and-governance-to-the-employee-endpoint-302890925.html). ## How is agent-layer enforcement deployed? Common routes are a native hook the agent calls before each tool, a gateway or proxy that mediates tool and model traffic, a platform integration with a SaaS or cloud agent service, and a sandbox that runs the agent. Noma describes enforcement for agents including Claude Code, Cursor, Codex and Windsurf on employee machines. The agent layer is the one most vendors are building toward in 2026, so documentation changes fast; read dates matter. ## Which tools document it, read 16 Nov 2026 On 16 Nov 2026, 14 of 16 tools had enforcement at the agent layer documented specifically, 2 in general terms or partly, and 0 none found. Levels follow the method page: a tool earns yes only when a vendor page or document describes this specifically, and the cell notes quote what the evidence says. - Documented specifically: Aim Security, Bay, Check Point, Cisco, CrowdStrike, HiddenLayer, Lasso Security, Microsoft, Noma Security, Operant AI, Pillar Security, SentinelOne (Prompt Security), Straiker and Zenity. - Documented in general terms or in part: Cloudflare and Palo Alto Networks. Sources: [Aim Security evidence page for this cell](https://www.catonetworks.com/platform/agentic-ai-security/); [Bay evidence page for this cell](https://www.bay.io/agent-runtime-enforcement?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=asset-dim); [Check Point evidence page for this cell](https://www.checkpoint.com/ai-security/ai-agent-security/). ## Related pages - [The full Enforcement Point Map](/enforcement-point-map/) - [Application layer](/enforcement-point-map/application-layer/) - [Endpoint layer](/enforcement-point-map/endpoint-layer/) - [Method](/enforcement-point-map/methodology/) - [Vendor profiles](/vendors/) --- # What Does Endpoint Layer AI Runtime Security Enforce? URL: https://airuntimesecuritytools.com/enforcement-point-map/endpoint-layer/ Updated: 2026-11-16 Endpoint-layer enforcement acts on the machine itself: processes, files, packages and connections an AI agent causes. On 16 Nov 2026, 9 of 16 tools documented it specifically, 6 in general terms or in part, and 1 had none found. It deploys as an endpoint agent or sensor. ## Endpoint layer: every vendor **Endpoint layer by vendor** | Vendor | Level | Note | Evidence | |---|---|---|---| | [Aim Security](/vendors/aim-security/) | Yes | Endpoint client blocks tool calls in supported local agents. | [aim-security-40](/vendors/aim-security/#fact-aim-security-40), [aim-security-41](/vendors/aim-security/#fact-aim-security-41) | | [Bay](/vendors/bay/) | Yes | Enforcer binary decides on the machine in under 4ms. | [bay-46](/vendors/bay/#fact-bay-46), [bay-48](/vendors/bay/#fact-bay-48), [bay-49](/vendors/bay/#fact-bay-49) | | [Cisco](/vendors/cisco/) | Yes | Hooks and sandbox on developer machines. | [cisco-11](/vendors/cisco/#fact-cisco-11), [cisco-12](/vendors/cisco/#fact-cisco-12), [cisco-13](/vendors/cisco/#fact-cisco-13) | | [CrowdStrike](/vendors/crowdstrike/) | Yes | Falcon sensor acts on processes, commands and packages on the endpoint. | [crowdstrike-09](/vendors/crowdstrike/#fact-crowdstrike-09), [crowdstrike-16](/vendors/crowdstrike/#fact-crowdstrike-16) | | [Microsoft](/vendors/microsoft/) | Yes | Defender for Endpoint agent runtime protection, preview. | [microsoft-19](/vendors/microsoft/#fact-microsoft-19), [microsoft-20](/vendors/microsoft/#fact-microsoft-20) | | [Noma Security](/vendors/noma-security/) | Yes | Endpoint agents covered with hooks. | [noma-security-08](/vendors/noma-security/#fact-noma-security-08), [noma-security-09](/vendors/noma-security/#fact-noma-security-09) | | [Operant AI](/vendors/operant-ai/) | Yes | Endpoint Protector enforces on the device. | [operant-ai-09](/vendors/operant-ai/#fact-operant-ai-09), [operant-ai-27](/vendors/operant-ai/#fact-operant-ai-27) | | [Pillar Security](/vendors/pillar-security/) | Yes | Pre-action controls on developer machines. | [pillar-security-09](/vendors/pillar-security/#fact-pillar-security-09), [pillar-security-10](/vendors/pillar-security/#fact-pillar-security-10) | | [Zenity](/vendors/zenity/) | Yes | Native hooks on coding agents on developer machines. | [zenity-50](/vendors/zenity/#fact-zenity-50), [zenity-20](/vendors/zenity/#fact-zenity-20) | | [Check Point](/vendors/check-point/) | Partial | Desktop agent for Workforce AI; machine-level action control not documented. | [check-point-08](/vendors/check-point/#fact-check-point-08), [check-point-10](/vendors/check-point/#fact-check-point-10) | | [HiddenLayer](/vendors/hiddenlayer/) | Partial | On-device hooks gate coding-agent actions for Claude Code, Cursor and Copilot CLI only, not general machine actions. | [hiddenlayer-10](/vendors/hiddenlayer/#fact-hiddenlayer-10), [hiddenlayer-11](/vendors/hiddenlayer/#fact-hiddenlayer-11), [hiddenlayer-20](/vendors/hiddenlayer/#fact-hiddenlayer-20) | | [Lasso Security](/vendors/lasso-security/) | Partial | Says enforcement runs at the execution layer on the device, but no endpoint install detail published. | [lasso-security-28](/vendors/lasso-security/#fact-lasso-security-28), [lasso-security-26](/vendors/lasso-security/#fact-lasso-security-26) | | [Palo Alto Networks](/vendors/palo-alto-networks/) | Partial | Only community example hooks; endpoint enforcement sits in a separate profile. | [palo-alto-networks-50](/vendors/palo-alto-networks/#fact-palo-alto-networks-50), [palo-alto-networks-69](/vendors/palo-alto-networks/#fact-palo-alto-networks-69) | | [SentinelOne (Prompt Security)](/vendors/sentinelone/) | Partial | Endpoint AI agent described in vendor blog; machine actions not itemized. | [sentinelone-32](/vendors/sentinelone/#fact-sentinelone-32), [sentinelone-13](/vendors/sentinelone/#fact-sentinelone-13) | | [Straiker](/vendors/straiker/) | Partial | Thin client for coding agents; general endpoint enforcement not documented. | [straiker-07](/vendors/straiker/#fact-straiker-07), [straiker-06](/vendors/straiker/#fact-straiker-06) | | [Cloudflare](/vendors/cloudflare/) | None found | Nothing found on endpoint actions for these AI products. | [cloudflare-24](/vendors/cloudflare/#fact-cloudflare-24) | ## What does an enforcement point at the endpoint layer see? An endpoint control sees actions as the operating system sees them: processes started, files touched, packages arriving and connections opened. CrowdStrike says the Falcon sensor gives OS-level visibility into agent processes, commands and resource access. Because it watches the machine, it also sees what the agent's own tools do not report, such as code the agent wrote and then ran. ## What does it miss? The operating system does not see meaning. Operant's argument is that the kernel sees a permitted system call and the network sees an approved endpoint while the damage is done through technically authorized actions, which is why it pairs intent checks with the lower layers. Endpoint coverage also ends at managed devices. An agent running in a cloud service or a SaaS platform is outside it. Sources: [Operant: Introducing Operant Semantic Firewall](https://www.operant.ai/art-kubed/introducing-operant-semantic-firewall). ## How is endpoint-layer enforcement deployed? Deployment is by an endpoint agent or sensor, usually pushed through existing device management or an endpoint security console, sometimes with a browser extension for web AI use. Bay says it installs through an existing MDM or EDR in under five minutes, as reported by SiliconANGLE on 10 Nov 2026. Endpoint agents work with agent-layer hooks: the hook decides on a tool call, and the sensor can stop the resulting action at the operating system. ## Which tools document it, read 16 Nov 2026 On 16 Nov 2026, 9 of 16 tools had enforcement at the endpoint layer documented specifically, 6 in general terms or partly, and 1 none found. Levels follow the method page: a tool earns yes only when a vendor page or document describes this specifically, and the cell notes quote what the evidence says. - Documented specifically: Aim Security, Bay, Cisco, CrowdStrike, Microsoft, Noma Security, Operant AI, Pillar Security and Zenity. - Documented in general terms or in part: Check Point, HiddenLayer, Lasso Security, Palo Alto Networks, SentinelOne (Prompt Security) and Straiker. - None found on 16 Nov 2026: Cloudflare. Sources: [Aim Security evidence page for this cell](https://www.catonetworks.com/solutions/ai-security/); [Bay evidence page for this cell](https://siliconangle.com/2026/11/10/ai-security-startup-bay-launches-with-50m-to-vet-agent-actions-before-they-run/); [Cisco evidence page for this cell](https://cisco-ai-defense.github.io/defenseclaw/docs/support-matrix/). ## Related pages - [The full Enforcement Point Map](/enforcement-point-map/) - [Agent layer](/enforcement-point-map/agent-layer/) - [Model layer](/enforcement-point-map/model-layer/) - [Method](/enforcement-point-map/methodology/) - [Vendor profiles](/vendors/) --- # AI runtime security datasets URL: https://airuntimesecuritytools.com/datasets/ Updated: 2026-11-16 Two datasets are free to download as CSV and JSON under CC BY 4.0: the Runtime Action Type Matrix, covering six action types across 16 vendors, and the Enforcement Point Map, covering six enforcement points. Both rest on the 16 Nov 2026 evidence files. ## Downloads - [Runtime Action Type Matrix](/datasets/runtime-action-type-matrix/): [CSV](/datasets/runtime-action-type-matrix.csv), [JSON](/datasets/runtime-action-type-matrix.json) - [Enforcement Point Map](/datasets/enforcement-point-map/): [CSV](/datasets/enforcement-point-map.csv), [JSON](/datasets/enforcement-point-map.json) ## License and citation CC BY 4.0. AI Runtime Security Tools, Runtime Action Type Matrix and Enforcement Point Map, edition 16 Nov 2026. https://airuntimesecuritytools.com/. Licensed CC BY 4.0. ## Related pages - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) - [Method](/method/) - [The ranking](/ranking/) --- # Runtime Action Type Matrix dataset: CSV and JSON download URL: https://airuntimesecuritytools.com/datasets/runtime-action-type-matrix/ Updated: 2026-11-16 The Runtime Action Type Matrix is published as a CSV file and a JSON file under CC BY 4.0. It holds 128 vendor and dimension cells across 8 dimensions, read on 2026-11-16. Cite it with the line below when you reuse the data, and link back to this page so readers can check the files against the sources. ## Files - [Runtime Action Type Matrix as CSV](/datasets/runtime-action-type-matrix.csv) - [Runtime Action Type Matrix as JSON](/datasets/runtime-action-type-matrix.json) ## How to cite AI Runtime Security Tools, Runtime Action Type Matrix and Enforcement Point Map, edition 16 Nov 2026. https://airuntimesecuritytools.com/. Licensed CC BY 4.0. ## Related pages - [The Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Method](/runtime-action-type-matrix/methodology/) - [All datasets](/datasets/) - [Scoring method](/method/) --- # Enforcement Point Map dataset: CSV and JSON download URL: https://airuntimesecuritytools.com/datasets/enforcement-point-map/ Updated: 2026-11-16 The Enforcement Point Map is published as a CSV file and a JSON file under CC BY 4.0. It holds 64 vendor and dimension cells across 4 dimensions, read on 2026-11-16. Cite it with the line below when you reuse the data, and link back to this page so readers can check the files against the sources. ## Files - [Enforcement Point Map as CSV](/datasets/enforcement-point-map.csv) - [Enforcement Point Map as JSON](/datasets/enforcement-point-map.json) ## How to cite AI Runtime Security Tools, Runtime Action Type Matrix and Enforcement Point Map, edition 16 Nov 2026. https://airuntimesecuritytools.com/. Licensed CC BY 4.0. ## Related pages - [The Enforcement Point Map](/enforcement-point-map/) - [Method](/enforcement-point-map/methodology/) - [All datasets](/datasets/) - [Scoring method](/method/) --- # AI runtime security guides URL: https://airuntimesecuritytools.com/guides/ Updated: 2026-11-16 Twelve guides explain AI runtime security for practitioners: what it is, how AIDR differs from EDR, inline versus out-of-band enforcement, failure modes, latency budgets, logging, RAG pipelines, metrics, cloud provider controls and open-source projects. Each claim carries a source and a 16 Nov 2026 read date. ## 12 guides - [AI Runtime Security Explained](/guides/ai-runtime-security-explained/): What AI runtime security is, which actions it stops, where it enforces and how it differs from posture management, with sources read on 16 Nov 2026. - [AI Runtime Security vs Guardrails vs Red Teaming](/guides/ai-runtime-security-vs-guardrails-vs-red-teaming/): How AI runtime security, guardrails and red teaming differ: when each acts, what each produces and how they fit together, with sources. - [AIDR vs EDR vs XDR: What Each Watches](/guides/aidr-vs-edr-vs-xdr/): What AI detection and response (AIDR), EDR and XDR each watch, and how to run them together, with vendor documentation read on 16 Nov 2026. - [Cloud Provider Built-In Runtime Controls](/guides/cloud-provider-built-in-runtime-controls/): What Amazon Bedrock Guardrails, Azure AI Content Safety and Google Model Armor include, and when a separate AI runtime security tool is needed. - [Fail Open or Fail Closed: What Vendors Publish](/guides/fail-open-or-fail-closed-what-vendors-publish/): What happens when an AI runtime control fails, which vendors document a failure mode, and what to ask for, using documents read on 16 Nov 2026. - [Inline or Out-of-Band Runtime Enforcement](/guides/inline-or-out-of-band-runtime-enforcement/): Inline enforcement can stop an AI action before it completes; out-of-band monitoring sees it afterward. How to choose, with vendor and cloud sources. - [Open-Source AI Runtime Security Projects](/guides/open-source-ai-runtime-security-projects/): Open-source projects that enforce at runtime: NeMo Guardrails, Llama Guard, Prompt Guard, LLM Guard and Guardrails AI, compared with commercial tools. - [Runtime Protection for Customer-Facing LLM Apps](/guides/runtime-protection-for-customer-facing-llm-apps/): Attacks on a public LLM app and the controls that sit in the request path: input checks, output checks, data filters and logs, with sources. - [Runtime Security for RAG Pipelines](/guides/runtime-security-for-rag-pipelines/): What goes wrong in a retrieval-augmented generation pipeline at runtime and where to place controls: ingest, retrieval, prompt, output. - [Runtime Security Latency Budgets for LLM Apps](/guides/runtime-security-latency-budgets-for-llm-apps/): How much delay an AI runtime control may add to an LLM app, how vendors state it, and how to measure it yourself, with sources read on 16 Nov 2026. - [Runtime Security Metrics: False Positives and Blocks](/guides/runtime-security-metrics-false-positives-and-blocked-rate/): Which metrics show whether an AI runtime control works: blocked rate, false positives, misses and latency, and how to set a false positive budget. - [What to Log From Runtime Security Decisions](/guides/what-to-log-from-runtime-security-decisions/): Which AI runtime security decisions to record, what fields to keep, and how to send them to a SIEM, with vendor documentation read on 16 Nov 2026. ## Related pages - [Glossary](/glossary/) - [The ranking](/ranking/) - [FAQ](/faq/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # AI runtime security explained URL: https://airuntimesecuritytools.com/guides/ai-runtime-security-explained/ Updated: 2026-11-16 AI runtime security stops a harmful step while an AI system runs: a prompt, a model answer, an agent tool call or an action on a machine. It decides before the step completes. Posture management finds weaknesses in advance. Vendor documents read on 16 Nov 2026 show the two are separate jobs. ## What is AI runtime security? AI runtime security is a control that inspects an AI step as it happens and allows, alters or blocks it before it completes. The step can be a user prompt, a model response, an agent tool call or an action such as a shell command on a developer machine. Cisco describes the idea for AI Defense as guardrails in the network that block adversarial attacks and harmful responses in real time. HiddenLayer documents the same idea for coding agents: gates when a prompt is submitted, before a shell command or MCP tool runs, when tool output returns and when files are read or edited. Sources: [Cisco AI Defense](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html); [HiddenLayer Agent Harness Security architecture](https://docs.hiddenlayer.ai/docs/products/runtime/agent_harness/architecture.md). ## How does it differ from posture management? Posture management looks at configuration and exposure before or around a run, and runtime security acts during it. A posture tool can tell you an agent has broad file access. A runtime control is what stops the agent from reading a credentials file at the moment it tries. The two answer different questions, so many vendors sell both. This site ranks only the runtime job, and scores a vendor on what it stops before the action completes. Sources: [HiddenLayer Agent Harness Security architecture](https://docs.hiddenlayer.ai/docs/products/runtime/agent_harness/architecture.md). ## Which actions can it stop? The actions fall into two groups. Content actions are prompt injection, jailbreak attempts, sensitive data in prompts and outputs, and harmful model output. Machine actions are shell commands, file access, package installs, network calls and credential access. OWASP's 2025 list for LLM applications names the content risks as LLM01 prompt injection, LLM02 sensitive information and LLM05 improper output handling, and names LLM06 excessive agency for agents given too much power. Few vendors document stops for every machine action; package install blocking is the rarest in the 16 Nov 2026 evidence. Sources: [OWASP Top 10 for LLM Applications 2025](https://genai.owasp.org/llm-top-10/). ## Where does a runtime control enforce? It can sit at the model layer, in front of an application, at the agent's tool calls or on the endpoint. Each point sees different things. A gateway sees traffic it is routed, and an endpoint hook sees the machine but may not see a SaaS agent. CrowdStrike says Falcon AI Gateway acts on routed AI traffic, and its agent-level hooks cover named coding agents. Palo Alto Networks documents a scan API that checks prompts and responses for any model. Neither claim covers the other's layer. Sources: [CrowdStrike Falcon AI Gateway](https://www.crowdstrike.com/en-us/blog/falcon-ai-gateway-one-control-point-for-ai-traffic/); [CrowdStrike advances Falcon Guardian](https://www.crowdstrike.com/en-us/blog/crowdstrike-advances-ai-security-and-safety-falcon-guardian/); [Prisma AIRS API Intercept overview](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/ai-runtime-security-api-intercept-overview.html). ## What should a buyer check first? Check that the vendor names the action it stops and says the block happens before completion. Then check for a published latency figure and a stated failure mode. Treat alert-only wording as detection, not a stop. MITRE ATLAS and the NIST AI 600-1 profile are useful for naming the risks you want covered before you ask vendors. Sources: [MITRE ATLAS](https://atlas.mitre.org/); [NIST AI 600-1, AI Risk Management Framework profile](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf). ## Why did this category appear now? Agents moved the harm from words to actions. A model that only answers can leak data or say something harmful. A model with tools can delete files, install packages and send data out, which OWASP's 2025 list calls excessive agency (LLM06). Vendors now document gates for these steps. HiddenLayer's docs describe blocking Cursor shell execution and file reads before they run, and Zenity says it can block or modify a dangerous tool call before it executes. Sources: [OWASP Top 10 for LLM Applications 2025](https://genai.owasp.org/llm-top-10/); [HiddenLayer Agent Harness Security architecture](https://docs.hiddenlayer.ai/docs/products/runtime/agent_harness/architecture.md). ## What does it not do? It does not remove the need to fix the underlying permissions, patch software or test a system before release. A control that stops a bad action is a last gate, and the first gates are least privilege and sound design. It also cannot stop what it cannot see. A gateway sees routed traffic only, a model-layer filter sees text only, and an endpoint hook sees only the agents it supports. Sources: [CrowdStrike Falcon AI Gateway](https://www.crowdstrike.com/en-us/blog/falcon-ai-gateway-one-control-point-for-ai-traffic/). ## Questions about this topic ### What is AI runtime security? AI runtime security is a control that inspects an AI step as it happens and allows, alters or blocks it before it completes. ## Sources - [OWASP Top 10 for LLM Applications 2025](https://genai.owasp.org/llm-top-10/) (genai.owasp.org), read 2026-11-16 - [MITRE ATLAS](https://atlas.mitre.org/) (atlas.mitre.org), read 2026-11-16 - [NIST AI 600-1, AI Risk Management Framework profile](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf) (nvlpubs.nist.gov), read 2026-11-16 - [Cisco AI Defense](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html) (cisco.com), read 2026-11-16 - [HiddenLayer Agent Harness Security architecture](https://docs.hiddenlayer.ai/docs/products/runtime/agent_harness/architecture.md) (docs.hiddenlayer.ai), read 2026-11-16 - [CrowdStrike Falcon AI Gateway](https://www.crowdstrike.com/en-us/blog/falcon-ai-gateway-one-control-point-for-ai-traffic/) (crowdstrike.com), read 2026-11-16 - [Prisma AIRS API Intercept overview](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/ai-runtime-security-api-intercept-overview.html) (docs.paloaltonetworks.com), read 2026-11-16 ## Related pages - [AI Runtime Security vs Guardrails vs Red Teaming](/guides/ai-runtime-security-vs-guardrails-vs-red-teaming/) - [Inline or Out-of-Band Runtime Enforcement](/guides/inline-or-out-of-band-runtime-enforcement/) - [Runtime Protection (glossary)](/glossary/runtime-protection/) - [All guides](/guides/) - [What to Log From Runtime Security Decisions](/guides/what-to-log-from-runtime-security-decisions/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # AI runtime security vs guardrails vs red teaming URL: https://airuntimesecuritytools.com/guides/ai-runtime-security-vs-guardrails-vs-red-teaming/ Updated: 2026-11-16 Guardrails are policies on what a model may say or do. Runtime security enforces policies and more while a system runs. Red teaming tests before launch and finds gaps. Sources read on 16 Nov 2026 show guardrails often sit inside runtime tools, and red teaming feeds both. ## Are guardrails the same as runtime security? Guardrails are one mechanism that runtime security uses, so they overlap but are not the same. A guardrail checks a prompt or response against a policy. Runtime security also covers agent tool calls, machine actions, logging and response. NVIDIA's NeMo Guardrails names five rail types: input, dialog, retrieval, execution and output. Input rails can reject or alter user input, and output rails can reject or alter what the model returns. Execution rails apply to the inputs and outputs of tools the model calls. Sources: [NVIDIA NeMo Guardrails](https://github.com/NVIDIA/NeMo-Guardrails). ## Where does red teaming fit? Red teaming fits before release and on a schedule, because it tests a system and reports what got through. It does not stop anything at runtime. Cisco shows the split in its own offer: AI Defense includes a free Explorer Edition described as red teaming, while the paid runtime protection places guardrails in the network. Findings from testing become runtime policy. Sources: [Cisco AI Defense](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html). ## What does each one produce? Guardrails produce policy decisions per request. Runtime security produces decisions plus a log of allowed and blocked actions. Red teaming produces a report of attack paths that succeeded. MITRE ATLAS gives a shared vocabulary of adversary tactics against AI systems, useful for mapping red team findings to runtime rules. Zenity says its detections map to OWASP LLM and MITRE ATLAS. Sources: [MITRE ATLAS](https://atlas.mitre.org/); [Zenity AI detection and response](https://zenity.io/platform/ai-detection-and-response). ## Who is responsible for each control? Guardrails often sit with the application team, runtime security with the security team and red teaming with either. Pick one accountable team per control, because a guardrail changed by an application release can silently weaken a security rule. A cloud provider's guardrail, such as Amazon Bedrock Guardrails, can be applied to a model call or called on its own through an API without invoking a model, which lets security review the policy separately. Sources: [Amazon Bedrock Guardrails](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html). ## How should you sequence them? Start with red teaming to learn what your system is exposed to, put guardrails on the model path, then add runtime enforcement where actions have side effects. NIST AI 600-1, the AI risk profile, treats these as risk management actions across the lifecycle, not one product. Re-test after each policy change. Sources: [NIST AI 600-1, AI Risk Management Framework profile](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf). ## What do the three miss alone? Guardrails alone miss agent actions, because a text policy cannot see a shell command that runs on a machine. Runtime security alone has no view of weaknesses nobody has tried yet. Red teaming alone gives a snapshot that ages with each model or prompt change. Treat them as a loop: test, set policy, enforce, log, then test again with what the logs showed. Sources: [NVIDIA NeMo Guardrails](https://github.com/NVIDIA/NeMo-Guardrails); [Cisco AI Defense](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html). ## What should you ask a vendor about each? Ask what policies come built in, what you can write yourself, and whether findings from testing can be loaded as runtime rules. Ask whether the product only detects or also blocks. Where the answer is not public, expect the score to reflect that: a claim without a public source is capped by this site's method. ## How do you keep them consistent? Keep one written policy that all three read: what the system may say, what it may do and what it may touch. Guardrail rules, runtime rules and red team test cases should each trace to a line in that policy. When a red team finds a gap, change the policy first, then the rules, then re-test. A fix made only in a guardrail can leave the agent action path open. ## Sources - [NVIDIA NeMo Guardrails](https://github.com/NVIDIA/NeMo-Guardrails) (github.com), read 2026-11-16 - [Cisco AI Defense](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html) (cisco.com), read 2026-11-16 - [MITRE ATLAS](https://atlas.mitre.org/) (atlas.mitre.org), read 2026-11-16 - [Zenity AI detection and response](https://zenity.io/platform/ai-detection-and-response) (zenity.io), read 2026-11-16 - [Amazon Bedrock Guardrails](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html) (docs.aws.amazon.com), read 2026-11-16 - [NIST AI 600-1, AI Risk Management Framework profile](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf) (nvlpubs.nist.gov), read 2026-11-16 ## Related pages - [AI Runtime Security Explained](/guides/ai-runtime-security-explained/) - [Open-Source AI Runtime Security Projects](/guides/open-source-ai-runtime-security-projects/) - [Block vs Detect (glossary)](/glossary/block-vs-detect/) - [All guides](/guides/) - [AIDR vs EDR vs XDR: What Each Watches](/guides/aidr-vs-edr-vs-xdr/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # AIDR vs EDR vs XDR: what each one watches URL: https://airuntimesecuritytools.com/guides/aidr-vs-edr-vs-xdr/ Updated: 2026-11-16 EDR watches processes, files and network activity on endpoints. XDR correlates signals across endpoint, identity, cloud and email. AIDR watches AI interactions: prompts, responses, tool calls and agent steps. CrowdStrike, Zenity and Noma all use the AIDR label as of 16 Nov 2026. ## What is AI detection and response? AI detection and response (AIDR) is detection and response applied to AI activity: prompts, model outputs, tool calls and agent steps, with the option to block. Vendors use the label differently, so read what is being watched. Zenity says its AIDR tracks each agent step, including triggers, retrievals, tool calls and handoffs, and blocks prompt injection before the agent acts on it. Noma says its AI-DR policies can block an action, mask data, steer the agent, bring in a human or trigger a killswitch. Sources: [Zenity AI detection and response](https://zenity.io/platform/ai-detection-and-response); [Noma AI-DR](https://noma.security/products/ai-dr). ## How does AIDR differ from EDR? EDR sees what a program does on a machine, and AIDR sees what an AI system is asked and does. An EDR sensor sees that a coding agent spawned a shell. It may not know a prompt injected into a document caused it. CrowdStrike says Falcon Guardian is the successor to Falcon AI Detection and Response and discovers AI agents on Windows, macOS and Linux endpoints, and its deployment notes describe the Falcon sensor on endpoints, the same sensor family used for EDR. Sources: [CrowdStrike on Falcon Guardian](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/). ## Where does XDR fit? XDR joins signals from several sources into incidents. AIDR events are one more source to join. Microsoft says alerts from Defender's agent runtime protection flow into Defender XDR incidents and agent data is queryable in advanced hunting. Check whether the AIDR tool exports to your SIEM by name. HiddenLayer documents outcomes sent to a webhook, Splunk or Azure Sentinel. Sources: [Microsoft Defender AI agent runtime protection overview](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-runtime-protection-overview); [HiddenLayer policy outcomes](https://docs.hiddenlayer.ai/docs/products/runtime/agentic/policy_outcomes.md). ## Do you need both? Most teams running coding agents or customer-facing LLM apps need both, because the two watch different layers. EDR is not designed to read a prompt, and AIDR is not designed to replace malware detection. Where one vendor sells both, the question is whether the AI controls block before execution or only alert. CrowdStrike says agent-level hooks for Claude Code, Codex, GitHub Copilot, OpenCode and Cursor block malicious activity before harmful actions run. Sources: [CrowdStrike advances Falcon Guardian](https://www.crowdstrike.com/en-us/blog/crowdstrike-advances-ai-security-and-safety-falcon-guardian/). ## What should the response path look like? Plan response before enabling blocking. Decide who sees a blocked action, what a user is told and which cases go to a person. Microsoft lists preview limits for its endpoint runtime protection, including that Claude Code may keep processing a tool response in block mode after a post-response block. Read vendor limits and test your own agents. Sources: [Microsoft Defender AI agent runtime protection overview](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-runtime-protection-overview). ## Which AI actions does EDR miss? EDR is built around processes and files, so it can miss a model call that leaks data, a prompt that changes agent behavior or a tool call to a remote service that looks like normal traffic. Microsoft documents that its network inspection for agent runtime protection skips agents using certificate pinning or HTTP/3, which shows how coverage depends on how the agent talks. Sources: [Microsoft Defender AI agent runtime protection overview](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-runtime-protection-overview). ## Which is not a replacement for the other? Neither replaces the other. EDR and XDR still carry malware and intrusion detection. AIDR adds the AI layer. Vendors that merge them, such as CrowdStrike with its Falcon sensor base, reduce the number of agents on a machine but still need AI-specific policy. Push Security-style identity detection and response is a separate topic not covered here. Sources: [CrowdStrike on Falcon Guardian](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/). ## Questions about this topic ### What is AI detection and response? AI detection and response is detection and response applied to AI activity: prompts, model outputs, tool calls and agent steps. ### Is AIDR the same as EDR? No. EDR watches processes, files and network activity on endpoints, while AIDR watches AI interactions. ## Sources - [Zenity AI detection and response](https://zenity.io/platform/ai-detection-and-response) (zenity.io), read 2026-11-16 - [Noma AI-DR](https://noma.security/products/ai-dr) (noma.security), read 2026-11-16 - [CrowdStrike on Falcon Guardian](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/) (crowdstrike.com), read 2026-11-16 - [CrowdStrike advances Falcon Guardian](https://www.crowdstrike.com/en-us/blog/crowdstrike-advances-ai-security-and-safety-falcon-guardian/) (crowdstrike.com), read 2026-11-16 - [Microsoft Defender AI agent runtime protection overview](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-runtime-protection-overview) (learn.microsoft.com), read 2026-11-16 - [HiddenLayer policy outcomes](https://docs.hiddenlayer.ai/docs/products/runtime/agentic/policy_outcomes.md) (docs.hiddenlayer.ai), read 2026-11-16 ## Related pages - [AI Runtime Security Explained](/guides/ai-runtime-security-explained/) - [What to Log From Runtime Security Decisions](/guides/what-to-log-from-runtime-security-decisions/) - [All guides](/guides/) - [AI Runtime Security vs Guardrails vs Red Teaming](/guides/ai-runtime-security-vs-guardrails-vs-red-teaming/) - [Cloud Provider Built-In Runtime Controls](/guides/cloud-provider-built-in-runtime-controls/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # Cloud provider built-in runtime controls URL: https://airuntimesecuritytools.com/guides/cloud-provider-built-in-runtime-controls/ Updated: 2026-11-16 Amazon Bedrock Guardrails, Azure AI Content Safety with Prompt Shields and Google Model Armor each screen prompts and responses for harmful content, sensitive data and prompt attacks. None of the three pages read on 16 Nov 2026 describes blocking shell commands on a developer machine. ## What do Azure, AWS and Google Cloud include? Each provider ships a content and attack filter for model inputs and outputs. AWS documents Bedrock Guardrails with content filters including a Prompt Attack category, denied topics, word filters and sensitive information filters that block or mask personal data. All three are managed services you call over an API, so they add a network hop and a dependency to each request, which is worth measuring before launch. Azure documents Prompt Shields, which detect user prompt attacks and document attacks. Google documents Model Armor, which screens prompts and responses for prompt injection, jailbreaks, sensitive data and harmful content. Sources: [Amazon Bedrock Guardrails components](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-components.html); [Prompt Shields in Azure AI Content Safety](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection); [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview). ## How do they enforce? Each can be called by your application. Bedrock Guardrails can be applied during a model call or invoked alone through the ApplyGuardrail API without calling a model. Model Armor offers an inspect-only mode and an inspect-and-block mode, so you can test before blocking. Microsoft says Prompt Shields in Foundry can block inputs classed as attacks before the model processes them. Sources: [Amazon Bedrock Guardrails](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html); [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview); [Prompt Shields in Azure AI Content Safety](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection). ## Do they work outside their own cloud? Some do. Google says Model Armor can be used whether you deploy AI in Google Cloud or on other cloud providers. Bedrock's ApplyGuardrail API takes text from any source. Check the data path: text leaves your application to be screened, so review residency and logging. Sources: [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview); [Amazon Bedrock Guardrails](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html). ## When is a separate tool needed? Add a separate tool when the harm is an agent action, not text: a shell command, a file read, a package install or a network call. The three provider pages read on 16 Nov 2026 are about prompts and responses. Also add one when you need endpoint coverage, a cross-cloud audit trail in one place, or enforcement in agents that run outside the provider. Sources: [Amazon Bedrock Guardrails components](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-components.html); [What is Azure AI Content Safety](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/overview); [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview). ## How do you combine them? Use the provider control as the base layer for model calls in that cloud, then add agent and endpoint controls where actions happen. Keep one owner for the combined policy so rules do not conflict. Microsoft also documents Defender for Endpoint runtime protection for agents, currently public preview, which shows a provider can extend past text. Sources: [What's new in Microsoft Defender for Endpoint](https://learn.microsoft.com/en-us/defender-endpoint/whats-new-in-microsoft-defender-endpoint). ## What do the pages not say? The pages read on 16 Nov 2026 give no latency figure in milliseconds for these controls on the pages sampled, and the failure behavior for a service outage is something to confirm in the service terms or by testing. Ask your cloud account team for both before relying on the control in a customer path. Sources: [Amazon Bedrock Guardrails](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html); [What is Azure AI Content Safety](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/overview). ## What does a rollout look like? Create a policy template, run it in log-only mode on a copy of traffic, review hits, then enforce. Google's Model Armor documentation suggests separate templates for input and output, so traceability and tuning stay simple. Record the policy version with each decision. Sources: [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview). ## What should you check in each service? Check regions, quotas, pricing and how logs are stored. Check whether the filter supports your languages and whether sensitive text is stored by the service. Review the provider's documentation again each quarter, since these services change often. Sources: [Amazon Bedrock Guardrails](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html); [What is Azure AI Content Safety](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/overview); [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview). ## Sources - [Amazon Bedrock Guardrails](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html) (docs.aws.amazon.com), read 2026-11-16 - [Amazon Bedrock Guardrails components](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-components.html) (docs.aws.amazon.com), read 2026-11-16 - [Prompt Shields in Azure AI Content Safety](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection) (learn.microsoft.com), read 2026-11-16 - [What is Azure AI Content Safety](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/overview) (learn.microsoft.com), read 2026-11-16 - [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview) (cloud.google.com), read 2026-11-16 - [What's new in Microsoft Defender for Endpoint](https://learn.microsoft.com/en-us/defender-endpoint/whats-new-in-microsoft-defender-endpoint) (learn.microsoft.com), read 2026-11-16 ## Related pages - [Runtime Protection for Customer-Facing LLM Apps](/guides/runtime-protection-for-customer-facing-llm-apps/) - [Open-Source AI Runtime Security Projects](/guides/open-source-ai-runtime-security-projects/) - [All guides](/guides/) - [AIDR vs EDR vs XDR: What Each Watches](/guides/aidr-vs-edr-vs-xdr/) - [Fail Open or Fail Closed: What Vendors Publish](/guides/fail-open-or-fail-closed-what-vendors-publish/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # Fail open or fail closed: what vendors publish URL: https://airuntimesecuritytools.com/guides/fail-open-or-fail-closed-what-vendors-publish/ Updated: 2026-11-16 Failure mode decides whether an AI action is blocked or allowed when the control breaks. Of 16 vendors read on 16 Nov 2026, Cisco documents a fail-closed default with an operator override, and HiddenLayer documents a timeout setting for one detector. Most others have none found. ## What happens when the control fails? When a runtime control fails or times out, the system either blocks the action (fail closed) or lets it through (fail open). The choice trades safety against availability, and it should be a decision you made, not a default you found later. Cisco's Kong plugin for AI Defense fails closed by default: traffic is blocked when inspection is unavailable. An operator can set it to fail open. Sources: [Cisco AI Defense Kong plugin](https://github.com/cisco-ai-defense/ai-defense-kong-plugin). ## Which vendors document their failure mode? On 16 Nov 2026 the documented cases are few. Cisco publishes its default. HiddenLayer's documentation describes a fail-open or fail-closed timeout for its code detector, but none found for a platform-wide default. For Bay, Check Point, CrowdStrike, Lasso, Straiker, Zenity and others, the evidence files record fail-open or fail-closed behavior as none found. That is not a statement about the product, only about what is public. Sources: [Cisco AI Defense Kong plugin](https://github.com/cisco-ai-defense/ai-defense-kong-plugin); [HiddenLayer runtime release notes](https://docs.hiddenlayer.ai/docs/products/runtime/releasenotes.md). ## What about timeouts? A timeout is the point where a slow control counts as failed. Palo Alto Networks documents a verdict timeout setting, but none found for the action taken when it expires. Ask each vendor for the timeout value, the default action on expiry and whether you can set the default per policy or per action. Sources: [Prisma AIRS API Intercept overview](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/ai-runtime-security-api-intercept-overview.html); [Palo Alto Networks: integrate Anthropic inference hooks](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/integrate-anthropic-inference-hooks.html). ## How should you choose a failure mode? Choose per action, not per product. Fail closed for actions that cannot be undone, such as credential access or data leaving the company. Fail open, with an alert, for low-risk content checks in a customer path where an outage would stop sales. Write the choice in the runbook and test it by cutting the control's connection in a staging environment. ## What can you ask a vendor? Ask four questions: what is the default on failure, can I change it, what is the timeout, and what do I see when it happens. Request the answer in public documentation, since a claim without a public source can score at most 7 in this site's method. On-device decisions avoid a network dependency. Bay says its decisions are made locally with no cloud round trip, though its fail-open or fail-closed behavior is none found. Sources: [Bay agent runtime enforcement](https://www.bay.io/agent-runtime-enforcement?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=guide). ## Why does this matter for AI more than for other tools? An AI control sits in the path of work that runs continuously and unattended. A fail-closed control that stops can halt a pipeline overnight, and a fail-open control that stops silently removes protection while looking healthy. Agents act without a person watching each step, so a quiet failure is the worse case. Alert on the failure itself, not only on blocked actions. ## How do you test the failure path? Block the control's network path or stop its process in a staging environment, send test actions of each type and record what the application does. Repeat with a slow response to check the timeout. Write down what you saw next to the vendor's documentation. A gap between the two is a finding for the vendor, and a reason to set the default yourself. ## What does a good public statement look like? A good statement names the default, the timeout, the way to change each and the log entry written when the control fails. Cisco's plugin README covers the default and the override. Where a statement is missing, record none found on the read date and ask the vendor to publish it. Sources: [Cisco AI Defense Kong plugin](https://github.com/cisco-ai-defense/ai-defense-kong-plugin). ## Questions about this topic ### What happens when the control fails? The system either blocks the action (fail closed) or lets it through (fail open). Cisco documents a fail-closed default for its Kong plugin, with an operator override. ## Sources - [Cisco AI Defense Kong plugin](https://github.com/cisco-ai-defense/ai-defense-kong-plugin) (github.com), read 2026-11-16 - [HiddenLayer runtime release notes](https://docs.hiddenlayer.ai/docs/products/runtime/releasenotes.md) (docs.hiddenlayer.ai), read 2026-11-16 - [Prisma AIRS API Intercept overview](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/ai-runtime-security-api-intercept-overview.html) (docs.paloaltonetworks.com), read 2026-11-16 - [Bay agent runtime enforcement](https://www.bay.io/agent-runtime-enforcement?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=guide) (bay.io), read 2026-11-16 ## Related pages - [Runtime Security Latency Budgets for LLM Apps](/guides/runtime-security-latency-budgets-for-llm-apps/) - [Fail Open (glossary)](/glossary/fail-open/) - [Fail Closed (glossary)](/glossary/fail-closed/) - [All guides](/guides/) - [Cloud Provider Built-In Runtime Controls](/guides/cloud-provider-built-in-runtime-controls/) - [Inline or Out-of-Band Runtime Enforcement](/guides/inline-or-out-of-band-runtime-enforcement/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # Inline or out-of-band runtime enforcement URL: https://airuntimesecuritytools.com/guides/inline-or-out-of-band-runtime-enforcement/ Updated: 2026-11-16 Inline enforcement sits in the path of an AI action and can stop it before it completes. Out-of-band monitoring reads a copy and can only alert or respond afterward. Google Model Armor and Palo Alto Networks document both modes as of 16 Nov 2026. ## What is the difference between inline and out-of-band? An inline control sits in the path of the action, so nothing proceeds until it answers. An out-of-band control reads a copy of the traffic or a log, so the action has usually completed by the time it reports. Google's Model Armor shows both in one product: an inspect-only mode that logs violations without stopping the request, and an inspect-and-block mode that prevents the offending content from proceeding. Sources: [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview). ## Which can stop an action? Only inline control can stop an action before it completes. Out-of-band tools can alert, open a ticket or revoke access, which limits damage but does not prevent the first step. Palo Alto Networks documents that a Block action in an AI security profile intercepts and stops the request and records a violation. Its alert action records without stopping. Sources: [Prisma AIRS AI security profile actions](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/configure-allow-alert-block-actions-ai-security-profiles.html). ## What does inline cost? Inline adds latency to every action and makes the control a dependency. If it fails, the system either blocks work or lets it through. Cisco's Kong gateway plugin fails closed by default, so traffic is blocked when inspection is unavailable, and an operator can set it to fail open. Few vendors state this default publicly. Sources: [Cisco AI Defense Kong plugin](https://github.com/cisco-ai-defense/ai-defense-kong-plugin). ## Where do inline controls sit? They sit in a gateway or proxy, in an SDK call inside the application, in a hook the agent runs before a tool call, or in an endpoint sensor. Cloudflare's AI Gateway DLP, on a Block outcome, discards the provider response and returns a status 400 error. HiddenLayer documents gates in coding agents before shell commands and MCP tools run. Sources: [Cloudflare AI Gateway DLP](https://developers.cloudflare.com/ai-gateway/features/dlp/); [HiddenLayer Agent Harness Security architecture](https://docs.hiddenlayer.ai/docs/products/runtime/agent_harness/architecture.md). ## When is out-of-band enough? Out-of-band fits low-risk actions, a first rollout where you need to measure violation rates, and systems where a delay would harm users. Model Armor's documentation says inspect-only mode is useful for testing new policies and auditing without affecting live traffic. Move to inline for actions with side effects you cannot undo, such as deleting data, sending credentials or running installed code. Sources: [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview). ## How should you roll out blocking? Start in a log-only mode for a week or two, review what would have been blocked, then turn blocking on for one action type at a time. Keep an override path for staff whose work is wrongly stopped. Cisco's DefenseClaw documentation lists four outcomes, allow, alert, ask and block, which lets you apply a softer outcome before a hard one. Sources: [DefenseClaw guardrail docs](https://cisco-ai-defense.github.io/defenseclaw/docs/guardrail/). ## Does the choice change by layer? Yes. At the model layer, inline means the response waits for an output check. At the agent layer, inline means a hook runs before a tool call. At the endpoint, inline means a sensor decides before a process proceeds. Microsoft's Defender runtime protection for agents has block and audit modes, and the block mode is public preview with listed limits. Sources: [Microsoft Defender AI agent runtime protection overview](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-runtime-protection-overview). ## What are the common mistakes? The first mistake is calling a copy-of-traffic tool a blocker. If the vendor page says detects or alerts, it did not stop the action. The second is running inline with no tested failure mode, which turns an outage into a business risk. The third is enforcing everywhere at once. Roll out by action type, so a false positive is easy to trace to one rule. ## Sources - [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview) (cloud.google.com), read 2026-11-16 - [Prisma AIRS AI security profile actions](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/configure-allow-alert-block-actions-ai-security-profiles.html) (docs.paloaltonetworks.com), read 2026-11-16 - [Cisco AI Defense Kong plugin](https://github.com/cisco-ai-defense/ai-defense-kong-plugin) (github.com), read 2026-11-16 - [Cloudflare AI Gateway DLP](https://developers.cloudflare.com/ai-gateway/features/dlp/) (developers.cloudflare.com), read 2026-11-16 - [HiddenLayer Agent Harness Security architecture](https://docs.hiddenlayer.ai/docs/products/runtime/agent_harness/architecture.md) (docs.hiddenlayer.ai), read 2026-11-16 ## Related pages - [Fail Open or Fail Closed: What Vendors Publish](/guides/fail-open-or-fail-closed-what-vendors-publish/) - [Inline Enforcement (glossary)](/glossary/inline-enforcement/) - [Block vs Detect (glossary)](/glossary/block-vs-detect/) - [All guides](/guides/) - [Open-Source AI Runtime Security Projects](/guides/open-source-ai-runtime-security-projects/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # Open-source AI runtime security projects URL: https://airuntimesecuritytools.com/guides/open-source-ai-runtime-security-projects/ Updated: 2026-11-16 Open-source projects can screen prompts and outputs at runtime: NVIDIA NeMo Guardrails, Meta's Llama Guard and Prompt Guard, LLM Guard and Guardrails AI. They give control and no vendor fee, but you run them. Repositories were read on 16 Nov 2026. ## Which open-source projects enforce at runtime? Four widely used projects each enforce at the model layer. NVIDIA NeMo Guardrails adds programmable rails to LLM applications, with input, dialog, retrieval, execution and output rails. Meta's Purple Llama repository includes Llama Guard, Llama Prompt Guard 2 and LlamaFirewall. Each is a library or model you integrate yourself, so enforcement happens only where your code calls it, and any path that skips the call is unprotected. LLM Guard from Protect AI lists prompt and output scanners, such as prompt injection, secrets and sensitive-data scanners. Guardrails AI is a Python framework that runs input and output guards and a hub of pre-built checks. Sources: [NVIDIA NeMo Guardrails](https://github.com/NVIDIA/NeMo-Guardrails); [Meta Purple Llama](https://github.com/meta-llama/PurpleLlama); [LLM Guard](https://github.com/protectai/llm-guard); [Guardrails AI](https://github.com/guardrails-ai/guardrails). ## What does each model do? Llama Guard 3 is a suite of models for safety classification of text and images. Llama Prompt Guard 2 detects prompt injection and jailbreaking attacks and is offered in two sizes. These are classifiers, so your code must act on their verdicts. They do not run as a gateway or on your endpoint by themselves. Sources: [Llama Guard 3](https://github.com/meta-llama/PurpleLlama/tree/main/Llama-Guard3); [Llama Prompt Guard 2](https://www.llama.com/docs/model-cards-and-prompt-formats/prompt-guard/). ## How do they compare with commercial tools? Open source gives code you can read and change. Commercial tools add hosted inference, agent hooks, endpoint coverage, logs and support. The open-source projects here screen text, and none of the pages read describes blocking a shell command or package install on a machine. Commercial vendors differ widely. Cisco publishes Apache-2.0 code for parts of its AI Defense tooling, and HiddenLayer publishes open-source code and release notes. Sources: [Cisco AI Defense](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html); [HiddenLayer runtime release notes](https://docs.hiddenlayer.ai/docs/products/runtime/releasenotes.md). ## What does it cost to run them? You pay in engineering time and compute: hosting models, tuning thresholds, updating rules and building logs. Guardrails AI announced on 6 Jul 2026 that its validators are moving to standard PyPI packages and its hosted remote inferencing would end, with a planned cutoff of 25 Aug 2026. That kind of change is a maintenance cost to plan for. Sources: [Guardrails AI](https://github.com/guardrails-ai/guardrails). ## When should you start with open source? Start with open source to learn your traffic, prototype policies, or keep data in your environment. Move to a commercial tool where you need endpoint enforcement, managed updates, support and an audit trail. Whichever you choose, measure false positives and latency on your own prompts first. ## What should you check before adopting one? Check the license, the last release date, how detections are updated, and whether the project documents how to run it at your volume. Check what happens on an error, since an open-source library inherits your application's failure mode. Run it against your own test set and record false positives before enforcing. Sources: [NVIDIA NeMo Guardrails](https://github.com/NVIDIA/NeMo-Guardrails); [LLM Guard](https://github.com/protectai/llm-guard). ## Which fits which job? Use Llama Guard to classify content safety, Prompt Guard to flag prompt attacks, NeMo Guardrails to shape dialog and tool use, LLM Guard for scanner checks and Guardrails AI for structured output validation. None of these replaces an agent-action control. They are building blocks around the model. Sources: [Llama Guard 3](https://github.com/meta-llama/PurpleLlama/tree/main/Llama-Guard3); [Llama Prompt Guard 2](https://www.llama.com/docs/model-cards-and-prompt-formats/prompt-guard/); [NVIDIA NeMo Guardrails](https://github.com/NVIDIA/NeMo-Guardrails); [LLM Guard](https://github.com/protectai/llm-guard); [Guardrails AI](https://github.com/guardrails-ai/guardrails). ## How do you keep them current? Pin versions, watch the repository's release notes and re-run your test set after every upgrade. Guardrails AI's notice about moving validators to PyPI packages shows how a project can change its distribution, so read release notes before an upgrade. Assign one person to own updates, since no vendor will. Sources: [Guardrails AI](https://github.com/guardrails-ai/guardrails). ## Questions about this topic ### Is there open-source AI runtime protection? Yes. NeMo Guardrails, Llama Guard, Llama Prompt Guard 2, LLM Guard and Guardrails AI screen prompts and outputs, though you host and tune them. ## Sources - [NVIDIA NeMo Guardrails](https://github.com/NVIDIA/NeMo-Guardrails) (github.com), read 2026-11-16 - [Meta Purple Llama](https://github.com/meta-llama/PurpleLlama) (github.com), read 2026-11-16 - [Llama Guard 3](https://github.com/meta-llama/PurpleLlama/tree/main/Llama-Guard3) (github.com), read 2026-11-16 - [Llama Prompt Guard 2](https://www.llama.com/docs/model-cards-and-prompt-formats/prompt-guard/) (llama.com), read 2026-11-16 - [LLM Guard](https://github.com/protectai/llm-guard) (github.com), read 2026-11-16 - [Guardrails AI](https://github.com/guardrails-ai/guardrails) (github.com), read 2026-11-16 - [Cisco AI Defense](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html) (cisco.com), read 2026-11-16 - [HiddenLayer runtime release notes](https://docs.hiddenlayer.ai/docs/products/runtime/releasenotes.md) (docs.hiddenlayer.ai), read 2026-11-16 ## Related pages - [Cloud Provider Built-In Runtime Controls](/guides/cloud-provider-built-in-runtime-controls/) - [AI Runtime Security vs Guardrails vs Red Teaming](/guides/ai-runtime-security-vs-guardrails-vs-red-teaming/) - [All guides](/guides/) - [Inline or Out-of-Band Runtime Enforcement](/guides/inline-or-out-of-band-runtime-enforcement/) - [Runtime Protection for Customer-Facing LLM Apps](/guides/runtime-protection-for-customer-facing-llm-apps/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # Runtime protection for customer-facing LLM apps URL: https://airuntimesecuritytools.com/guides/runtime-protection-for-customer-facing-llm-apps/ Updated: 2026-11-16 A public LLM app faces prompt injection, sensitive data exposure, harmful output and runaway use, all named in OWASP's 2025 list. Controls sit in the request path: input checks, output checks and data filters. Sources read on 16 Nov 2026 show cloud and edge vendors document these checks. ## What attacks hit a public LLM app? Anyone who can type can try to subvert the model, so the first attacks are prompt injection, attempts to extract the system prompt, coaxing out sensitive data and abusing cost. OWASP's 2025 list names LLM01 prompt injection, LLM02 sensitive information, LLM07 system prompt leakage and LLM10 unbounded consumption. Output is a risk too. OWASP names LLM05 improper output handling for responses passed on without checks. Sources: [OWASP Top 10 for LLM Applications 2025](https://genai.owasp.org/llm-top-10/). ## Which controls sit in the request path? Three checks sit in the path: an input check before the model, an output check before the user, and a data filter on both. Amazon Bedrock Guardrails evaluates input prompts and model completions against defined filters, and can also be called on its own through the ApplyGuardrail API. Azure's Prompt Shields analyze text for user prompt attacks and document attacks, and in Foundry can block flagged inputs before the model processes them. Sources: [Amazon Bedrock Guardrails](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html); [Prompt Shields in Azure AI Content Safety](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection). ## What can an edge or gateway layer do? It can score and block before traffic reaches your servers. Cloudflare says AI Security for Apps scores each prompt for injection so customers can block with WAF rules, and AI Gateway Guardrails can flag or block prompts before the model. Google's Model Armor screens prompts and responses and also works for applications on other clouds. Sources: [Cloudflare AI Security for Apps](https://developers.cloudflare.com/waf/detections/ai-security-for-apps/prompt-injection/); [Cloudflare AI Gateway Guardrails](https://developers.cloudflare.com/ai-gateway/features/guardrails/); [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview). ## How do you handle sensitive data? Redact or block at the boundary. Bedrock's sensitive information filters can block or mask personal data in inputs and model responses, based on probabilistic detection. Probabilistic means some misses and some false blocks, so log what is masked and sample it. Sources: [Amazon Bedrock Guardrails components](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-components.html). ## What should you log and limit? Log each decision with the policy that fired, and set limits on tokens and request rates so one user cannot run up cost. Model Armor's documentation warns that false positives can degrade the user experience by blocking legitimate prompts, so tune thresholds on real traffic before enforcing. Sources: [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview). ## What does a minimum setup look like? A minimum setup has an input check for prompt attacks, an output check for harmful content and sensitive data, rate and token limits, and a log of decisions. Most cloud providers cover the first two with a built-in filter. Then add monitoring for the system prompt and tools. OWASP's LLM07 covers system prompt leakage, and its LLM06 covers tools given too much power. Sources: [OWASP Top 10 for LLM Applications 2025](https://genai.owasp.org/llm-top-10/); [What is Azure AI Content Safety](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/overview). ## What should you not rely on? Do not rely on the system prompt as a security boundary, on a single filter, or on a vendor claim with no public method. Google describes false positive trade-offs for its filter settings, so each filter needs tuning on real traffic. Layer an input check, an output check and a limit on what tools the app can call, and review blocked and allowed samples every week. Sources: [Prompt Shields in Azure AI Content Safety](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection); [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview). ## Which events should page a person? Page a person for a burst of blocked prompt attacks from one source, a spike in sensitive data blocks, a control failure and any change to the policy outside a release. Send the rest to a daily review. Keep the thresholds in the runbook, and test the alert path before launch. ## Sources - [OWASP Top 10 for LLM Applications 2025](https://genai.owasp.org/llm-top-10/) (genai.owasp.org), read 2026-11-16 - [Amazon Bedrock Guardrails](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html) (docs.aws.amazon.com), read 2026-11-16 - [Amazon Bedrock Guardrails components](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-components.html) (docs.aws.amazon.com), read 2026-11-16 - [Prompt Shields in Azure AI Content Safety](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection) (learn.microsoft.com), read 2026-11-16 - [Cloudflare AI Security for Apps](https://developers.cloudflare.com/waf/detections/ai-security-for-apps/prompt-injection/) (developers.cloudflare.com), read 2026-11-16 - [Cloudflare AI Gateway Guardrails](https://developers.cloudflare.com/ai-gateway/features/guardrails/) (developers.cloudflare.com), read 2026-11-16 - [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview) (cloud.google.com), read 2026-11-16 ## Related pages - [Runtime Security Latency Budgets for LLM Apps](/guides/runtime-security-latency-budgets-for-llm-apps/) - [Cloud Provider Built-In Runtime Controls](/guides/cloud-provider-built-in-runtime-controls/) - [Input Filtering (glossary)](/glossary/input-filtering/) - [All guides](/guides/) - [Open-Source AI Runtime Security Projects](/guides/open-source-ai-runtime-security-projects/) - [Runtime Security for RAG Pipelines](/guides/runtime-security-for-rag-pipelines/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # Runtime security for retrieval-augmented generation pipelines URL: https://airuntimesecuritytools.com/guides/runtime-security-for-rag-pipelines/ Updated: 2026-11-16 A retrieval-augmented generation pipeline pulls outside text into the prompt, so a poisoned document can steer the model. Controls belong at retrieval, at the prompt and at the output. Sources read on 16 Nov 2026 include OWASP's vector weaknesses entry and NeMo Guardrails retrieval rails. ## What can go wrong in a RAG pipeline at runtime? Retrieved text enters the prompt as if it were trusted, so a document can carry instructions that change what the model does. Stored vectors can also expose data to users who should not see it. The same pattern holds for any pipeline that places text from a search index, a ticket system or a file share into a prompt, so treat all such text as untrusted input. OWASP's 2025 list names LLM08 vector and embedding weaknesses and LLM01 prompt injection, which includes instructions hidden in retrieved content. Sources: [OWASP Top 10 for LLM Applications 2025](https://genai.owasp.org/llm-top-10/). ## Where should a control sit? Put a control at each hop: on the retrieved chunks, on the assembled prompt and on the answer. NeMo Guardrails has retrieval rails for this, where a rail can reject a chunk and keep it from prompting the model, or alter it to mask sensitive data. Azure's Prompt Shields include document attack detection, which analyzes supplied documents for attack patterns separately from the user's prompt. Sources: [NVIDIA NeMo Guardrails](https://github.com/NVIDIA/NeMo-Guardrails); [Prompt Shields in Azure AI Content Safety](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection). ## How do you limit what retrieval can return? Enforce the user's access rights at retrieval time, so the index returns only chunks the user may read. Filtering after the model answers is too late for data already in the prompt. Add a data filter on the answer as well. Bedrock's sensitive information filters can block or mask personal data in model responses. Sources: [Amazon Bedrock Guardrails components](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-components.html). ## What about agents that retrieve? An agent that retrieves and then calls tools turns a bad document into an action. Authorize the tool call separately from the retrieval, as NeMo's execution rails do for tools the model calls. HiddenLayer documents gates when tool output returns in coding agents, which is the same idea for content coming back from a tool. Sources: [NVIDIA NeMo Guardrails](https://github.com/NVIDIA/NeMo-Guardrails); [HiddenLayer Agent Harness Security architecture](https://docs.hiddenlayer.ai/docs/products/runtime/agent_harness/architecture.md). ## How do you test it? Seed a test index with harmless marker text that tries to change the model's behavior, and confirm each control catches it. Record which hop caught it and which did not. Repeat after any change to the index, the prompt template or the guardrail policy. ## What do the vendors say? Vendor coverage of retrieval is thin. Cloudflare's AI Gateway DLP scans prompts and responses and, on a block, discards the provider response, and Palo Alto Networks documents a scan API that checks prompts and model responses for any model. Neither page describes checking retrieved chunks as a named step, so ask vendors whether their check sees retrieved text. Sources: [Cloudflare AI Gateway DLP](https://developers.cloudflare.com/ai-gateway/features/dlp/); [Prisma AIRS API Intercept overview](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/ai-runtime-security-api-intercept-overview.html). ## What is the minimum for a first rollout? Start with access checks at retrieval, a prompt attack check on the assembled prompt and an output filter. Log which documents were retrieved for each answer so a bad answer can be traced to its source. Add document scanning at ingest when you can, and re-scan on a schedule. Sources: [Prompt Shields in Azure AI Content Safety](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection); [Amazon Bedrock Guardrails components](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-components.html). ## Which signs show a poisoned source? Watch for answers that suddenly cite one document, for retrieved text containing instructions addressed to the model, and for tool calls that follow a retrieval with no user request. Log the source of each chunk so a suspect document can be removed and the affected answers found. Sources: [OWASP Top 10 for LLM Applications 2025](https://genai.owasp.org/llm-top-10/). ## Sources - [OWASP Top 10 for LLM Applications 2025](https://genai.owasp.org/llm-top-10/) (genai.owasp.org), read 2026-11-16 - [NVIDIA NeMo Guardrails](https://github.com/NVIDIA/NeMo-Guardrails) (github.com), read 2026-11-16 - [Prompt Shields in Azure AI Content Safety](https://learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/jailbreak-detection) (learn.microsoft.com), read 2026-11-16 - [Amazon Bedrock Guardrails components](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-components.html) (docs.aws.amazon.com), read 2026-11-16 - [HiddenLayer Agent Harness Security architecture](https://docs.hiddenlayer.ai/docs/products/runtime/agent_harness/architecture.md) (docs.hiddenlayer.ai), read 2026-11-16 ## Related pages - [Runtime Protection for Customer-Facing LLM Apps](/guides/runtime-protection-for-customer-facing-llm-apps/) - [Output Filtering (glossary)](/glossary/output-filtering/) - [Input Filtering (glossary)](/glossary/input-filtering/) - [All guides](/guides/) - [Runtime Security Latency Budgets for LLM Apps](/guides/runtime-security-latency-budgets-for-llm-apps/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # Runtime security latency budgets for LLM apps URL: https://airuntimesecuritytools.com/guides/runtime-security-latency-budgets-for-llm-apps/ Updated: 2026-11-16 A latency budget caps the time a runtime control may add to each request. Vendors state figures from under 4 ms on-device to under 300 ms in the request path, all as their own claims on 16 Nov 2026. Measure your own p95 before you commit. ## How much latency is acceptable? Set the budget from the user-facing target, then give the control a fixed share of it. A chat reply that streams may hide a short input check. A synchronous action such as a tool call can tolerate less delay. There is no standard figure. Published vendor claims differ by an order of magnitude: Check Point claims under 50 ms, Lasso says gateway enforcement completes in under 50 ms, and Straiker says Defend AI runs at under 300 ms. Sources: [Check Point AI Security](https://www.checkpoint.com/ai-security/); [Lasso AI gateway security](https://www.lasso.security/use-cases/ai-gateway-security); [Straiker Defend AI](https://www.straiker.ai/products/defend-ai). ## How do vendors measure it? Mostly they do not say. A figure on a product page rarely states the percentile, payload size or region. Operant's endpoint page shows a 0.7 ms P99 figure in a demo panel, which is a page display, not a stated service level. HiddenLayer's release notes say its Interactions API now returns detection latency, which lets a customer measure each detection instead of trusting a number. Sources: [Operant Endpoint Protector](https://www.operant.ai/platform/endpoint-protector); [HiddenLayer runtime release notes](https://docs.hiddenlayer.ai/docs/products/runtime/releasenotes.md). ## Do on-device checks change the budget? On-device checks remove the network trip, so the budget is mostly compute. Bay says endpoint decisions take under 4 ms with no cloud round trip. SiliconANGLE reported on 2 Sep 2026 that Lasso's CPU-only guardrail returns a verdict in under 5 ms, per Lasso's own benchmarks. Both are vendor-reported. Treat them as a starting point for a test. Sources: [Bay agent runtime enforcement](https://www.bay.io/agent-runtime-enforcement?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=guide); [SiliconANGLE on Lasso LEAP, 2 Sep 2026](https://siliconangle.com/2026/09/02/lasso-security-launches-cpu-only-ai-guardrail-raises-30m/). ## What adds delay besides the control itself? Network hops to a SaaS inspection service, payload size, number of checks chained, and whether the response is held until the output check ends. Cloudflare documents buffering behavior for streamed responses in its guardrails, which matters because holding a stream for an output check delays the first token. Sources: [Cloudflare AI Gateway Guardrails](https://developers.cloudflare.com/ai-gateway/features/guardrails/). ## How do you test a vendor's number? Replay a sample of real prompts and tool calls, record p50, p95 and p99 with the control on and off, and repeat at peak load. Test the failure path too, by adding delay to the control and watching what your application does. Record the result with the date and version, because vendors change models and defaults. ## Which checks cost the most? Model-based classifiers cost more than pattern checks. Lasso's CPU-only guardrail claim shows vendors are working to shrink the model cost, and Meta's Prompt Guard 2 is offered in two sizes with the smaller one trading some attack prevention for speed. Choose the lightest check that meets your risk for each action, and reserve heavier checks for high-risk steps. Sources: [SiliconANGLE on Lasso LEAP, 2 Sep 2026](https://siliconangle.com/2026/09/02/lasso-security-launches-cpu-only-ai-guardrail-raises-30m/); [Llama Prompt Guard 2](https://www.llama.com/docs/model-cards-and-prompt-formats/prompt-guard/). ## What should a contract or proof of concept include? Include a latency target at p95 under your own load, the failure behavior in writing, and the right to re-test after each major release. Ask the vendor to show decision latency in the logs, as HiddenLayer documents for its Interactions API. Without logged latency you cannot hold anyone to a number. Sources: [HiddenLayer runtime release notes](https://docs.hiddenlayer.ai/docs/products/runtime/releasenotes.md). ## Does streaming change the answer? Streaming lets a user see text while the model still writes, so an output check has to either hold text back or check in chunks. Holding back adds delay to the first visible token, and checking in chunks can let a partial answer show before a block. Decide which risk you accept for each app, and test it with long answers, not only short ones. Sources: [Cloudflare AI Gateway Guardrails](https://developers.cloudflare.com/ai-gateway/features/guardrails/). ## Sources - [Check Point AI Security](https://www.checkpoint.com/ai-security/) (checkpoint.com), read 2026-11-16 - [Lasso AI gateway security](https://www.lasso.security/use-cases/ai-gateway-security) (lasso.security), read 2026-11-16 - [SiliconANGLE on Lasso LEAP, 2 Sep 2026](https://siliconangle.com/2026/09/02/lasso-security-launches-cpu-only-ai-guardrail-raises-30m/) (siliconangle.com), read 2026-11-16 - [Straiker Defend AI](https://www.straiker.ai/products/defend-ai) (straiker.ai), read 2026-11-16 - [Operant Endpoint Protector](https://www.operant.ai/platform/endpoint-protector) (operant.ai), read 2026-11-16 - [HiddenLayer runtime release notes](https://docs.hiddenlayer.ai/docs/products/runtime/releasenotes.md) (docs.hiddenlayer.ai), read 2026-11-16 - [Bay agent runtime enforcement](https://www.bay.io/agent-runtime-enforcement?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=guide) (bay.io), read 2026-11-16 - [Cloudflare AI Gateway Guardrails](https://developers.cloudflare.com/ai-gateway/features/guardrails/) (developers.cloudflare.com), read 2026-11-16 ## Related pages - [Fail Open or Fail Closed: What Vendors Publish](/guides/fail-open-or-fail-closed-what-vendors-publish/) - [Decision Latency (glossary)](/glossary/decision-latency/) - [Runtime Protection for Customer-Facing LLM Apps](/guides/runtime-protection-for-customer-facing-llm-apps/) - [All guides](/guides/) - [Runtime Security for RAG Pipelines](/guides/runtime-security-for-rag-pipelines/) - [Runtime Security Metrics: False Positives and Blocks](/guides/runtime-security-metrics-false-positives-and-blocked-rate/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # Runtime security metrics: false positives and blocked rate URL: https://airuntimesecuritytools.com/guides/runtime-security-metrics-false-positives-and-blocked-rate/ Updated: 2026-11-16 Measure a runtime control with five numbers: blocked rate, false positive rate, miss rate from testing, decision latency and coverage by action type. No standard benchmark exists, and vendors publish few figures. Sources read on 16 Nov 2026 include Google and Straiker documentation. ## How do you measure a runtime control? Measure outcomes by action type, not one overall score. Track how many actions of each type were allowed, blocked and asked, how long each decision took, and how often a person overrode a block. Coverage is a metric too: the share of agents, apps and machines that have the control on. A strong policy on half your estate is half a control. ## What is a good false positive rate? There is no published standard figure, so set a budget from the cost of a wrong block. A customer chat that blocks one real question in a hundred may be acceptable, and a build pipeline that blocks one command in a hundred may not. Google's Model Armor documentation says false positives can degrade the user experience by blocking legitimate prompts and responses, and that stricter confidence settings raise that risk. Straiker says Defend AI achieves 98.1 percent detection accuracy at under 300 ms, a vendor claim with no stated method. Sources: [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview); [Straiker Defend AI](https://www.straiker.ai/products/defend-ai). ## How do you find misses? Misses show up only through testing and incident review. Run a fixed set of safe test cases for each action type and record what got through, and review allowed actions near the policy edge. MITRE ATLAS and OWASP's list give categories to build test sets from. Keep test cases harmless and out of production data. Sources: [MITRE ATLAS](https://atlas.mitre.org/); [OWASP Top 10 for LLM Applications 2025](https://genai.owasp.org/llm-top-10/). ## How do audit and inspect-only modes help? Run new policies in a mode that logs without blocking, then compare what would have been blocked with what staff were doing. Model Armor's inspect-only mode exists for this: it logs violations without stopping the request. Move to blocking one action type at a time, and keep the log to show the change in blocked rate. Sources: [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview). ## How do you report to leadership? Report a short table: actions stopped by type, estimated false positives, decision latency at p95 and coverage. Add the open gaps in plain words. Avoid a single detection accuracy number unless the vendor publishes the test set and method, since without them it cannot be checked. ## What does blocked rate tell you? Blocked rate alone is ambiguous. A rising rate can mean more attacks, a policy that got stricter or a new app with odd traffic. Pair it with the reason for each block and a sample review. A falling rate after a policy change should prompt a check for misses, not celebration. ## How often should you review the numbers? Review weekly during rollout and monthly after. Look at the top blocked rules, the rules with the most overrides and the actions with no decisions at all, which can mean a gap in coverage. Change one rule at a time and keep the change log with the metric before and after. ## What should you ask vendors to publish? Ask for the test set, the method, the false positive rate on benign traffic and the date of the test. A detection figure with no benign-traffic rate is only half a measurement. This site scores latency and failure mode only as published, and the same standard applies to accuracy claims: no method, no credit. ## Sources - [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview) (cloud.google.com), read 2026-11-16 - [Straiker Defend AI](https://www.straiker.ai/products/defend-ai) (straiker.ai), read 2026-11-16 - [MITRE ATLAS](https://atlas.mitre.org/) (atlas.mitre.org), read 2026-11-16 - [OWASP Top 10 for LLM Applications 2025](https://genai.owasp.org/llm-top-10/) (genai.owasp.org), read 2026-11-16 ## Related pages - [What to Log From Runtime Security Decisions](/guides/what-to-log-from-runtime-security-decisions/) - [False Positive Budget (glossary)](/glossary/false-positive-budget/) - [Block vs Detect (glossary)](/glossary/block-vs-detect/) - [All guides](/guides/) - [Runtime Security Latency Budgets for LLM Apps](/guides/runtime-security-latency-budgets-for-llm-apps/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # What to log from runtime security decisions URL: https://airuntimesecuritytools.com/guides/what-to-log-from-runtime-security-decisions/ Updated: 2026-11-16 Log every decision, allowed and blocked, with the action, the policy that fired and the outcome. Then send the records to your SIEM. HiddenLayer, Cisco and Bay document exports as of 16 Nov 2026, in webhook, Splunk, Sentinel, OpenTelemetry and OCSF formats. ## Which decisions should be logged? Log all of them, not only blocks. A record of allowed actions shows what the control saw and lets you find a miss afterward, and a record of blocks shows what it stopped. Record the allow outcomes too, because a log of blocks alone cannot show a miss or measure a false positive rate later on. Cisco says every decision in AI Defense is recorded. DefenseClaw answers allow, alert, ask or block for a prompt, shell command or tool call, so each of the four outcomes needs a field. Sources: [DefenseClaw guardrail docs](https://cisco-ai-defense.github.io/defenseclaw/docs/guardrail/). ## What fields does a useful record hold? Keep the time, the user or agent, the action type, the target, the policy that fired, the decision and the reason. Add a session identifier so a person can replay what led up to a block. HiddenLayer documents recorded interactions and replayable sessions. Pillar says its activity view shows blocked decisions on session timelines. Sources: [HiddenLayer policy outcomes](https://docs.hiddenlayer.ai/docs/products/runtime/agentic/policy_outcomes.md); [Pillar Security agentic endpoint](https://www.pillar.security/platform/agentic-endpoint). ## How do you send them to a SIEM? Use the vendor's export in a format your SIEM reads, and confirm it carries the decision fields, not only alerts. OpenTelemetry defines a common log data model, and Cisco documents export to OpenTelemetry destinations. HiddenLayer documents delivery of detect, redact or block outcomes to a webhook, Splunk or Azure Sentinel. Bay says its events reach the SIEM in OCSF format. Sources: [OpenTelemetry logs](https://opentelemetry.io/docs/concepts/signals/logs/); [HiddenLayer policy outcomes](https://docs.hiddenlayer.ai/docs/products/runtime/agentic/policy_outcomes.md); [Bay agent runtime enforcement](https://www.bay.io/agent-runtime-enforcement?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=guide); [Bay AI agent investigation](https://www.bay.io/ai-agent-investigation?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=guide); [Cisco: Protect enterprise coding agents with DefenseClaw and NVIDIA OpenShell](https://blogs.cisco.com/ai/protect-enterprise-coding-agents-with-cisco-defenseclaw-and-nvidia-openshell). ## What should you avoid keeping? Avoid storing raw prompts and outputs longer than needed, since they hold the same sensitive data you are trying to protect. Cloudflare notes that its production mode logs only metadata. Keep the decision and a reference to content held in a restricted store with its own retention rule. Sources: [Cloudflare AI Gateway Guardrails](https://developers.cloudflare.com/ai-gateway/features/guardrails/). ## How do you use the logs? Review blocked and allowed counts by action type each week, sample allowed actions near the policy edge, and feed repeat blocks into response playbooks. Compare the log against the false positive budget you set, and keep the record for the audit period your regulator requires. ## Who needs the logs? Three groups read them: the security team for response, the application team for tuning and auditors for proof. Give each a view that fits and keep one source of truth. For regulated work, the ability to show what was blocked and why is the purpose of the record, so test retrieval of a month-old decision. ## What do vendors leave unclear? Many vendor pages do not name the SIEM or SOAR connectors. The evidence files for Check Point, Palo Alto Networks, Noma and SentinelOne record named SIEM export as none found on 16 Nov 2026, though some have a logs API. Ask for the field list and a sample export before you buy. Sources: [Check Point AI Security](https://www.checkpoint.com/ai-security/); [Prisma AIRS API Intercept overview](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/ai-runtime-security-api-intercept-overview.html). ## How long should records be kept? Keep them as long as your audit and incident rules require, and no longer for raw content. Decision metadata can be held for a long period at low risk, while prompt and output text should follow a shorter rule. Write the retention periods next to the field list so reviewers see both. ## Sources - [DefenseClaw guardrail docs](https://cisco-ai-defense.github.io/defenseclaw/docs/guardrail/) (cisco-ai-defense.github.io), read 2026-11-16 - [HiddenLayer policy outcomes](https://docs.hiddenlayer.ai/docs/products/runtime/agentic/policy_outcomes.md) (docs.hiddenlayer.ai), read 2026-11-16 - [Pillar Security agentic endpoint](https://www.pillar.security/platform/agentic-endpoint) (pillar.security), read 2026-11-16 - [OpenTelemetry logs](https://opentelemetry.io/docs/concepts/signals/logs/) (opentelemetry.io), read 2026-11-16 - [Bay agent runtime enforcement](https://www.bay.io/agent-runtime-enforcement?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=guide) (bay.io), read 2026-11-16 - [Cloudflare AI Gateway Guardrails](https://developers.cloudflare.com/ai-gateway/features/guardrails/) (developers.cloudflare.com), read 2026-11-16 ## Related pages - [Runtime Security Metrics: False Positives and Blocks](/guides/runtime-security-metrics-false-positives-and-blocked-rate/) - [SIEM Export (glossary)](/glossary/siem-export/) - [Telemetry (glossary)](/glossary/telemetry/) - [All guides](/guides/) - [AI Runtime Security Explained](/guides/ai-runtime-security-explained/) - [The ranking](/ranking/) - [Glossary](/glossary/) --- # AI runtime security glossary URL: https://airuntimesecuritytools.com/glossary/ Updated: 2026-11-16 Twenty-eight terms define the vocabulary of AI runtime security, from fail-open and fail-closed to decision latency, output filtering, tool-call authorization, reverse proxy, sidecar deployment and false positive budget. Each entry gives a definition and one sourced example read on 16 Nov 2026. ## 28 terms - [Anomaly Detection](/glossary/anomaly-detection/): Anomaly detection flags activity that departs from an expected pattern without matching a known bad signature. - [Behavioral Baseline](/glossary/behavioral-baseline/): A behavioral baseline is a learned picture of what an agent, or the person who runs it, normally does, such as which files it touches and which tools it calls. - [Block vs Detect](/glossary/block-vs-detect/): Block versus detect is the choice between stopping an action before it completes and only raising an alert. - [Canary Token](/glossary/canary-token/): A canary token is a fake credential, file or setting planted where nobody should look, which raises an alarm the moment it is touched. - [Decision Latency](/glossary/decision-latency/): Decision latency is the time an inline security check takes to return allow, block or redact for one prompt, reply or action. - [eBPF](/glossary/ebpf/): eBPF is a Linux kernel technology that runs small sandboxed programs inside the operating system to observe or control activity without changing kernel code. - [Fail Closed](/glossary/fail-closed/): Fail closed describes a control that blocks traffic or actions when the security check is unavailable, so a failure never opens a gap. - [Fail Open](/glossary/fail-open/): Fail open describes a control that lets traffic or actions through when the security check is unavailable, slow or times out. - [False Positive Budget](/glossary/false-positive-budget/): A false positive budget is the number of wrongly blocked or flagged safe actions a team agrees to tolerate in exchange for catching real attacks. - [Inline Enforcement](/glossary/inline-enforcement/): Inline enforcement means the security check sits in the path of the request, so the action waits for a verdict and can be blocked. - [Input Filtering](/glossary/input-filtering/): Input filtering inspects what is sent to a model, including user messages and retrieved documents, and can block or change it before the model sees it. - [Jailbreak](/glossary/jailbreak/): A jailbreak is an input crafted to make a model ignore its safety rules and produce content or take actions it was set up to refuse. - [Kernel-Level Monitoring](/glossary/kernel-level-monitoring/): Kernel-level monitoring watches the operating system's own events, such as system calls, process starts and file opens, from the most privileged part of the machine. - [Model Denial of Service](/glossary/model-denial-of-service/): Model denial of service is an attack or accident that overloads an AI model with requests or oversized inputs until the service slows, fails or runs up a bill. - [Output Filtering](/glossary/output-filtering/): Output filtering inspects what a model returns and can block, mask or rewrite it before a user or another system receives it. - [PII Redaction](/glossary/pii-redaction/): PII redaction replaces personal data in a prompt or reply with a placeholder or mask before it leaves the organization or reaches a user. - [Rate Limiting](/glossary/rate-limiting/): Rate limiting caps how many requests may reach a service in a time window and rejects the rest. - [Reverse Proxy](/glossary/reverse-proxy/): A reverse proxy for AI traffic is a gateway that sits in front of model providers and forwards requests after inspecting them. - [Runtime Protection](/glossary/runtime-protection/): Runtime protection is a security control that watches an AI system while it is running and can stop a harmful prompt, reply or action before it completes. - [SDK Integration](/glossary/sdk-integration/): An SDK integration adds the security check inside the application's own code through a vendor library that calls the service before and after model or tool calls. - [Sensitive Information Leakage](/glossary/sensitive-information-leakage/): Sensitive information leakage is the exposure of personal data, credentials, business secrets or model internals through an AI system's output. - [Sidecar Deployment](/glossary/sidecar-deployment/): A sidecar deployment runs the security check as a small companion process beside each application or agent, on the same host or pod. - [SIEM Export](/glossary/siem-export/): SIEM export is the sending of a runtime tool's alerts and decision records to a security information and event management system. - [SOAR Playbook](/glossary/soar-playbook/): A SOAR playbook is an automated set of response steps that runs when an alert or incident matches a rule. - [System Prompt Leakage](/glossary/system-prompt-leakage/): System prompt leakage is the exposure of the hidden instructions that steer a model, which can reveal rules, roles and any secrets placed in them. - [Telemetry](/glossary/telemetry/): Telemetry is the stream of records a runtime tool emits about what it saw and decided: prompts, tool calls, verdicts, timing and the policy that applied. - [Token Budget](/glossary/token-budget/): A token budget is a cap on how many model tokens a user, team, app or agent may consume in a period. - [Tool Call Authorization](/glossary/tool-call-authorization/): Tool call authorization is a decision, made before an AI agent runs a tool, on whether this agent may call this tool with these arguments. ## Related pages - [Guides](/guides/) - [FAQ](/faq/) - [The ranking](/ranking/) - [Runtime Action Type Matrix](/runtime-action-type-matrix/) - [Enforcement Point Map](/enforcement-point-map/) --- # Anomaly Detection: definition and example URL: https://airuntimesecuritytools.com/glossary/anomaly-detection/ Updated: 2026-11-16 Anomaly detection flags activity that departs from an expected pattern without matching a known bad signature. In AI runtime security it applies to sequences of prompts and tool calls as well as single events. It finds new attack forms but produces more uncertain alerts than rule matching. ## Example Straiker says Defend AI traces every user, model, tool-chain and agent-to-agent interaction to support root-cause analysis, anomaly detection and audit readiness. ## Sources - [Straiker page](https://www.straiker.ai/products/defend-ai) (straiker.ai), read 2026-11-16 ## Related pages - [Behavioral Baseline](/glossary/behavioral-baseline/) - [Telemetry](/glossary/telemetry/) - [False Positive Budget](/glossary/false-positive-budget/) - [All glossary terms](/glossary/) - [Tool Call Authorization](/glossary/tool-call-authorization/) - [Guides](/guides/) --- # Behavioral Baseline: definition and example URL: https://airuntimesecuritytools.com/glossary/behavioral-baseline/ Updated: 2026-11-16 A behavioral baseline is a learned picture of what an agent, or the person who runs it, normally does, such as which files it touches and which tools it calls. Deviations from it become alerts or triggers for review. A baseline needs time to learn and can raise false alarms while it does. ## Example CrowdStrike described on 4 Nov 2026 a Q4 roadmap feature in Next-Gen SIEM that learns the normal behavior of each AI agent and its owner, and marked it not yet available. ## Sources - [CrowdStrike page](https://www.crowdstrike.com/en-us/blog/crowdstrike-detects-and-stops-agent-insider-threats/) (crowdstrike.com), read 2026-11-16 ## Related pages - [Anomaly Detection](/glossary/anomaly-detection/) - [False Positive Budget](/glossary/false-positive-budget/) - [Telemetry](/glossary/telemetry/) - [All glossary terms](/glossary/) - [Block vs Detect](/glossary/block-vs-detect/) - [Guides](/guides/) --- # Block vs Detect: definition and example URL: https://airuntimesecuritytools.com/glossary/block-vs-detect/ Updated: 2026-11-16 Block versus detect is the choice between stopping an action before it completes and only raising an alert. Blocking prevents harm but risks stopping legitimate work. Detecting keeps work flowing but leaves the response to people or automation after the fact. ## Example Noma says it detects runtime behavior patterns that indicate credential theft, data exfiltration and AI memory tampering, and no block was stated for those detections. ## Sources - [Noma Security page](https://www.prnewswire.com/news-releases/noma-labs-discovers-critical-vulnerability-in-widely-adopted-open-source-ai-agent-platform-ruflo-302837494.html) (prnewswire.com), read 2026-11-16 ## Related pages - [Inline Enforcement](/glossary/inline-enforcement/) - [False Positive Budget](/glossary/false-positive-budget/) - [SOAR Playbook](/glossary/soar-playbook/) - [AI Runtime Security vs Guardrails vs Red Teaming](/guides/ai-runtime-security-vs-guardrails-vs-red-teaming/) - [Inline or Out-of-Band Runtime Enforcement](/guides/inline-or-out-of-band-runtime-enforcement/) - [Runtime Security Metrics: False Positives and Blocks](/guides/runtime-security-metrics-false-positives-and-blocked-rate/) - [All glossary terms](/glossary/) - [Behavioral Baseline](/glossary/behavioral-baseline/) - [Canary Token](/glossary/canary-token/) - [Guides](/guides/) --- # Canary Token: definition and example URL: https://airuntimesecuritytools.com/glossary/canary-token/ Updated: 2026-11-16 A canary token is a fake credential, file or setting planted where nobody should look, which raises an alarm the moment it is touched. In AI settings it can reveal an agent or attacker reading places they were not meant to reach. It detects misuse; it does not stop the first access. ## Example The Canarytokens guide lists, among its token types, an MCP Configuration Canarytoken and a Sensitive Command Canarytoken. ## Sources - [Canarytokens guide: Introduction](https://docs.canarytokens.org/guide/) (docs.canarytokens.org), read 2026-11-16 ## Related pages - [System Prompt Leakage](/glossary/system-prompt-leakage/) - [Block vs Detect](/glossary/block-vs-detect/) - [Telemetry](/glossary/telemetry/) - [All glossary terms](/glossary/) - [Decision Latency](/glossary/decision-latency/) - [Guides](/guides/) --- # Decision Latency: definition and example URL: https://airuntimesecuritytools.com/glossary/decision-latency/ Updated: 2026-11-16 Decision latency is the time an inline security check takes to return allow, block or redact for one prompt, reply or action. Because the check sits in the request path, the delay adds to every interaction. Vendors report it differently: as a median, a ceiling, or not at all. ## Example Lasso says its gateway enforcement, whether block, mask or alert, completes in under 50 ms. ## Sources - [Lasso Security page](https://www.lasso.security/use-cases/ai-gateway-security) (lasso.security), read 2026-11-16 ## Related pages - [Fail Open](/glossary/fail-open/) - [Inline Enforcement](/glossary/inline-enforcement/) - [False Positive Budget](/glossary/false-positive-budget/) - [Runtime Security Latency Budgets for LLM Apps](/guides/runtime-security-latency-budgets-for-llm-apps/) - [All glossary terms](/glossary/) - [Canary Token](/glossary/canary-token/) - [eBPF](/glossary/ebpf/) - [Guides](/guides/) --- # eBPF: definition and example URL: https://airuntimesecuritytools.com/glossary/ebpf/ Updated: 2026-11-16 eBPF is a Linux kernel technology that runs small sandboxed programs inside the operating system to observe or control activity without changing kernel code. Security tools use it to watch process, file and network events cheaply. It applies to Linux hosts and sees system calls, not the meaning of a prompt. ## Example The eBPF project describes the technology as running sandboxed programs in a privileged context such as the operating system kernel, extending it without changing kernel source code. ## Sources - [ebpf.io: What is eBPF?](https://ebpf.io/what-is-ebpf/) (ebpf.io), read 2026-11-16 ## Related pages - [Kernel-Level Monitoring](/glossary/kernel-level-monitoring/) - [Telemetry](/glossary/telemetry/) - [Inline Enforcement](/glossary/inline-enforcement/) - [All glossary terms](/glossary/) - [Decision Latency](/glossary/decision-latency/) - [Fail Closed](/glossary/fail-closed/) - [Guides](/guides/) --- # Fail Closed: definition and example URL: https://airuntimesecuritytools.com/glossary/fail-closed/ Updated: 2026-11-16 Fail closed describes a control that blocks traffic or actions when the security check is unavailable, so a failure never opens a gap. The cost is that an outage in the security service can stop the AI app too. Some products let the operator choose per deployment. ## Example Cisco's AI Defense Kong plugin fails closed by default, blocking traffic when inspection is unavailable, and an operator can set it to fail open. ## Sources - [Cisco page](https://github.com/cisco-ai-defense/ai-defense-kong-plugin) (github.com), read 2026-11-16 ## Related pages - [Fail Open](/glossary/fail-open/) - [Inline Enforcement](/glossary/inline-enforcement/) - [Decision Latency](/glossary/decision-latency/) - [Fail Open or Fail Closed: What Vendors Publish](/guides/fail-open-or-fail-closed-what-vendors-publish/) - [All glossary terms](/glossary/) - [eBPF](/glossary/ebpf/) - [Guides](/guides/) --- # Fail Open: definition and example URL: https://airuntimesecuritytools.com/glossary/fail-open/ Updated: 2026-11-16 Fail open describes a control that lets traffic or actions through when the security check is unavailable, slow or times out. It keeps the AI app running at the cost of leaving it unprotected during the outage. Whether a vendor publishes this behavior is a buying question. ## Example HiddenLayer's release notes say users can set a timeout for the code detector and choose whether it fails open or closed on timeout, and no platform-wide failure mode was found. ## Sources - [HiddenLayer page](https://docs.hiddenlayer.ai/docs/products/runtime/releasenotes.md) (docs.hiddenlayer.ai), read 2026-11-16 ## Related pages - [Fail Closed](/glossary/fail-closed/) - [Decision Latency](/glossary/decision-latency/) - [Inline Enforcement](/glossary/inline-enforcement/) - [Fail Open or Fail Closed: What Vendors Publish](/guides/fail-open-or-fail-closed-what-vendors-publish/) - [All glossary terms](/glossary/) - [False Positive Budget](/glossary/false-positive-budget/) - [Guides](/guides/) --- # False Positive Budget: definition and example URL: https://airuntimesecuritytools.com/glossary/false-positive-budget/ Updated: 2026-11-16 A false positive budget is the number of wrongly blocked or flagged safe actions a team agrees to tolerate in exchange for catching real attacks. It sets how aggressive a policy can be before people route around the control. Tools that report false positives and misses make the budget measurable. ## Example The Prisma AIRS page lists a November 2026 update adding false positive and false negative reporting and auto-remediation for API Intercept. ## Sources - [Palo Alto Networks page](https://www.paloaltonetworks.com/prisma/prisma-ai-runtime-security) (paloaltonetworks.com), read 2026-11-16 ## Related pages - [Block vs Detect](/glossary/block-vs-detect/) - [Decision Latency](/glossary/decision-latency/) - [Anomaly Detection](/glossary/anomaly-detection/) - [Runtime Security Metrics: False Positives and Blocks](/guides/runtime-security-metrics-false-positives-and-blocked-rate/) - [All glossary terms](/glossary/) - [Fail Open](/glossary/fail-open/) - [Inline Enforcement](/glossary/inline-enforcement/) - [Guides](/guides/) --- # Inline Enforcement: definition and example URL: https://airuntimesecuritytools.com/glossary/inline-enforcement/ Updated: 2026-11-16 Inline enforcement means the security check sits in the path of the request, so the action waits for a verdict and can be blocked. Out-of-band tools read a copy of the traffic or a log and can only alert. Inline checks add delay and create a failure-mode question, which is why latency and fail-open behavior are assessed apart. ## Example Cisco's DefenseClaw documentation says its guardrail checks a prompt, a shell command or a tool call before the agent acts and answers allow, alert, ask or block. ## Sources - [Cisco page](https://cisco-ai-defense.github.io/defenseclaw/docs/guardrail/) (cisco-ai-defense.github.io), read 2026-11-16 ## Related pages - [Fail Open](/glossary/fail-open/) - [Fail Closed](/glossary/fail-closed/) - [Block vs Detect](/glossary/block-vs-detect/) - [Inline or Out-of-Band Runtime Enforcement](/guides/inline-or-out-of-band-runtime-enforcement/) - [All glossary terms](/glossary/) - [False Positive Budget](/glossary/false-positive-budget/) - [Input Filtering](/glossary/input-filtering/) - [Guides](/guides/) --- # Input Filtering: definition and example URL: https://airuntimesecuritytools.com/glossary/input-filtering/ Updated: 2026-11-16 Input filtering inspects what is sent to a model, including user messages and retrieved documents, and can block or change it before the model sees it. It is the usual place to catch injected instructions, jailbreak attempts and sensitive data in prompts. A filter on text alone cannot see what an agent does afterward. ## Example Cloudflare says AI Gateway Guardrails work as a proxy between the application and model providers such as OpenAI and Anthropic. ## Sources - [Cloudflare page](https://developers.cloudflare.com/ai-gateway/features/guardrails/) (developers.cloudflare.com), read 2026-11-16 ## Related pages - [Output Filtering](/glossary/output-filtering/) - [Jailbreak](/glossary/jailbreak/) - [Reverse Proxy](/glossary/reverse-proxy/) - [Runtime Protection for Customer-Facing LLM Apps](/guides/runtime-protection-for-customer-facing-llm-apps/) - [Runtime Security for RAG Pipelines](/guides/runtime-security-for-rag-pipelines/) - [All glossary terms](/glossary/) - [Inline Enforcement](/glossary/inline-enforcement/) - [Guides](/guides/) --- # Jailbreak: definition and example URL: https://airuntimesecuritytools.com/glossary/jailbreak/ Updated: 2026-11-16 A jailbreak is an input crafted to make a model ignore its safety rules and produce content or take actions it was set up to refuse. It targets the model's own restrictions, unlike an injected instruction aimed at an app's data path. Runtime tools try to detect jailbreak attempts in inputs and unsafe results in outputs. ## Example Microsoft says Defender for Cloud AI threat protection works with Prompt Shields to alert on jailbreak, data leakage and credential theft. ## Sources - [Microsoft page](https://learn.microsoft.com/en-us/azure/defender-for-cloud/ai-threat-protection) (learn.microsoft.com), read 2026-11-16 ## Related pages - [Input Filtering](/glossary/input-filtering/) - [Output Filtering](/glossary/output-filtering/) - [Block vs Detect](/glossary/block-vs-detect/) - [All glossary terms](/glossary/) - [Kernel-Level Monitoring](/glossary/kernel-level-monitoring/) - [Guides](/guides/) --- # Kernel-Level Monitoring: definition and example URL: https://airuntimesecuritytools.com/glossary/kernel-level-monitoring/ Updated: 2026-11-16 Kernel-level monitoring watches the operating system's own events, such as system calls, process starts and file opens, from the most privileged part of the machine. It is hard for a user-space program to hide from. It reports what happened at the system level without knowing why an AI agent did it. ## Example Operant argues that the kernel sees a permitted system call and the network an approved endpoint while an agent's damage is done through technically authorized actions. ## Sources - [Operant AI: Introducing Operant Semantic Firewall](https://www.operant.ai/art-kubed/introducing-operant-semantic-firewall) (operant.ai), read 2026-11-16 ## Related pages - [eBPF](/glossary/ebpf/) - [Telemetry](/glossary/telemetry/) - [Behavioral Baseline](/glossary/behavioral-baseline/) - [All glossary terms](/glossary/) - [Jailbreak](/glossary/jailbreak/) - [Model Denial of Service](/glossary/model-denial-of-service/) - [Guides](/guides/) --- # Model Denial of Service: definition and example URL: https://airuntimesecuritytools.com/glossary/model-denial-of-service/ Updated: 2026-11-16 Model denial of service is an attack or accident that overloads an AI model with requests or oversized inputs until the service slows, fails or runs up a bill. OWASP now files it under unbounded consumption, which also covers denial of wallet. Rate limits and token budgets are the standard controls. ## Example OWASP's example of denial of wallet describes attackers starting a high volume of operations to exploit the cost-per-use model of cloud AI services. ## Sources - [OWASP LLM10:2025 Unbounded Consumption](https://genai.owasp.org/llmrisk/llm102025-unbounded-consumption/) (genai.owasp.org), read 2026-11-16 ## Related pages - [Rate Limiting](/glossary/rate-limiting/) - [Token Budget](/glossary/token-budget/) - [Telemetry](/glossary/telemetry/) - [All glossary terms](/glossary/) - [Kernel-Level Monitoring](/glossary/kernel-level-monitoring/) - [Output Filtering](/glossary/output-filtering/) - [Guides](/guides/) --- # Output Filtering: definition and example URL: https://airuntimesecuritytools.com/glossary/output-filtering/ Updated: 2026-11-16 Output filtering inspects what a model returns and can block, mask or rewrite it before a user or another system receives it. It targets leaked personal data, secrets, unsafe content and replies that break policy. It acts after the model has answered, so it does not change what the model was asked. ## Example Lasso says its gateway mode can block an interaction before a response returns, mask sensitive data in transit, or allow and alert. ## Sources - [Lasso Security page](https://www.lasso.security/use-cases/ai-gateway-security) (lasso.security), read 2026-11-16 ## Related pages - [Input Filtering](/glossary/input-filtering/) - [PII Redaction](/glossary/pii-redaction/) - [Sensitive Information Leakage](/glossary/sensitive-information-leakage/) - [Runtime Security for RAG Pipelines](/guides/runtime-security-for-rag-pipelines/) - [All glossary terms](/glossary/) - [Model Denial of Service](/glossary/model-denial-of-service/) - [Guides](/guides/) --- # PII Redaction: definition and example URL: https://airuntimesecuritytools.com/glossary/pii-redaction/ Updated: 2026-11-16 PII redaction replaces personal data in a prompt or reply with a placeholder or mask before it leaves the organization or reaches a user. It lets a request continue with the sensitive values removed instead of blocking it whole. Coverage varies by data type, file format and language. ## Example Check Point says Workforce AI Security protects data in prompts with inline redaction, and covers files and images with OCR-based redaction. ## Sources - [Check Point page](https://www.checkpoint.com/ai-security/workforce-ai-security/) (checkpoint.com), read 2026-11-16 ## Related pages - [Output Filtering](/glossary/output-filtering/) - [Sensitive Information Leakage](/glossary/sensitive-information-leakage/) - [Block vs Detect](/glossary/block-vs-detect/) - [All glossary terms](/glossary/) - [Rate Limiting](/glossary/rate-limiting/) - [Guides](/guides/) --- # Rate Limiting: definition and example URL: https://airuntimesecuritytools.com/glossary/rate-limiting/ Updated: 2026-11-16 Rate limiting caps how many requests may reach a service in a time window and rejects the rest. For AI apps it slows abusive callers and keeps costs predictable. Fixed windows reset on a clock; sliding windows count the most recent interval. ## Example Cloudflare's AI Gateway documentation gives the example of limiting an application to 100 requests per 60 seconds, with excess requests receiving a 429 Too Many Requests response. ## Sources - [Cloudflare AI Gateway docs: Rate limiting](https://developers.cloudflare.com/ai-gateway/features/rate-limiting/) (developers.cloudflare.com), read 2026-11-16 ## Related pages - [Token Budget](/glossary/token-budget/) - [Model Denial of Service](/glossary/model-denial-of-service/) - [Reverse Proxy](/glossary/reverse-proxy/) - [All glossary terms](/glossary/) - [PII Redaction](/glossary/pii-redaction/) - [Guides](/guides/) --- # Reverse Proxy: definition and example URL: https://airuntimesecuritytools.com/glossary/reverse-proxy/ Updated: 2026-11-16 A reverse proxy for AI traffic is a gateway that sits in front of model providers and forwards requests after inspecting them. Apps connect to it instead of to the provider, usually by changing a base URL. One proxy can cover many apps, but only the traffic routed through it. ## Example CrowdStrike says developers connect an existing SDK or coding tool to Falcon AI Gateway by changing the base URL and using a gateway key. ## Sources - [CrowdStrike page](https://www.crowdstrike.com/en-us/blog/falcon-ai-gateway-one-control-point-for-ai-traffic/) (crowdstrike.com), read 2026-11-16 ## Related pages - [SDK Integration](/glossary/sdk-integration/) - [Sidecar Deployment](/glossary/sidecar-deployment/) - [Input Filtering](/glossary/input-filtering/) - [All glossary terms](/glossary/) - [Rate Limiting](/glossary/rate-limiting/) - [Runtime Protection](/glossary/runtime-protection/) - [Guides](/guides/) --- # Runtime Protection: definition and example URL: https://airuntimesecuritytools.com/glossary/runtime-protection/ Updated: 2026-11-16 Runtime protection is a security control that watches an AI system while it is running and can stop a harmful prompt, reply or action before it completes. It differs from scanning a model before release or testing it with simulated attacks. The control may sit at the model, application, agent or endpoint layer. ## Example Microsoft says each runtime scan of an agent action is an inline check with minimal added latency, without publishing a millisecond figure. ## Sources - [Microsoft page](https://learn.microsoft.com/en-us/defender-endpoint/ai-agent-runtime-protection-overview) (learn.microsoft.com), read 2026-11-16 ## Related pages - [Inline Enforcement](/glossary/inline-enforcement/) - [Block vs Detect](/glossary/block-vs-detect/) - [Decision Latency](/glossary/decision-latency/) - [AI Runtime Security Explained](/guides/ai-runtime-security-explained/) - [All glossary terms](/glossary/) - [Reverse Proxy](/glossary/reverse-proxy/) - [SDK Integration](/glossary/sdk-integration/) - [Guides](/guides/) --- # SDK Integration: definition and example URL: https://airuntimesecuritytools.com/glossary/sdk-integration/ Updated: 2026-11-16 An SDK integration adds the security check inside the application's own code through a vendor library that calls the service before and after model or tool calls. It gives fine control and app context, but each app needs a code change. It covers only the apps that carry the library. ## Example The Cisco AI Defense Python SDK README says it can inspect AWS AgentCore and Strands event streams, releasing events only after server acknowledgement. ## Sources - [Cisco page](https://github.com/cisco-ai-defense/ai-defense-python-sdk) (github.com), read 2026-11-16 ## Related pages - [Reverse Proxy](/glossary/reverse-proxy/) - [Sidecar Deployment](/glossary/sidecar-deployment/) - [Tool Call Authorization](/glossary/tool-call-authorization/) - [All glossary terms](/glossary/) - [Runtime Protection](/glossary/runtime-protection/) - [Sensitive Information Leakage](/glossary/sensitive-information-leakage/) - [Guides](/guides/) --- # Sensitive Information Leakage: definition and example URL: https://airuntimesecuritytools.com/glossary/sensitive-information-leakage/ Updated: 2026-11-16 Sensitive information leakage is the exposure of personal data, credentials, business secrets or model internals through an AI system's output. OWASP lists it as its second risk for LLM applications. Controls combine data sanitization, output filtering and limits on what the model can reach. ## Example OWASP notes that restrictions placed in a system prompt about what data to return may not always be honored and could be bypassed via prompt injection. ## Sources - [OWASP LLM02:2025, sensitive information](https://genai.owasp.org/llmrisk/llm022025-sensitive-information-disclosure/) (genai.owasp.org), read 2026-11-16 ## Related pages - [PII Redaction](/glossary/pii-redaction/) - [Output Filtering](/glossary/output-filtering/) - [System Prompt Leakage](/glossary/system-prompt-leakage/) - [All glossary terms](/glossary/) - [SDK Integration](/glossary/sdk-integration/) - [Sidecar Deployment](/glossary/sidecar-deployment/) - [Guides](/guides/) --- # Sidecar Deployment: definition and example URL: https://airuntimesecuritytools.com/glossary/sidecar-deployment/ Updated: 2026-11-16 A sidecar deployment runs the security check as a small companion process beside each application or agent, on the same host or pod. It keeps the check close to the workload and avoids a network hop to a central service. It adds one more component to run and update per workload. ## Example A Check Point research post dated 15 Sep 2026 describes a sidecar control monitor that prevents harmful agent actions before they execute at sub-100ms latency. ## Sources - [Check Point page](https://blog.checkpoint.com/ai-security/synchronous-control-monitoring-preventing-harmful-agent-actions-in-real-time/) (blog.checkpoint.com), read 2026-11-16 ## Related pages - [Reverse Proxy](/glossary/reverse-proxy/) - [SDK Integration](/glossary/sdk-integration/) - [Decision Latency](/glossary/decision-latency/) - [All glossary terms](/glossary/) - [Sensitive Information Leakage](/glossary/sensitive-information-leakage/) - [SIEM Export](/glossary/siem-export/) - [Guides](/guides/) --- # SIEM Export: definition and example URL: https://airuntimesecuritytools.com/glossary/siem-export/ Updated: 2026-11-16 SIEM export is the sending of a runtime tool's alerts and decision records to a security information and event management system. It lets analysts correlate AI events with the rest of the estate and keep evidence under their own retention rules. Formats and fields differ by vendor. ## Example Bay says its events reach the customer's SIEM in OCSF format. ## Sources - [Bay page](https://www.bay.io/ai-agent-investigation?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=term) (bay.io), read 2026-11-16 ## Related pages - [Telemetry](/glossary/telemetry/) - [SOAR Playbook](/glossary/soar-playbook/) - [Anomaly Detection](/glossary/anomaly-detection/) - [What to Log From Runtime Security Decisions](/guides/what-to-log-from-runtime-security-decisions/) - [All glossary terms](/glossary/) - [Sidecar Deployment](/glossary/sidecar-deployment/) - [Guides](/guides/) --- # SOAR Playbook: definition and example URL: https://airuntimesecuritytools.com/glossary/soar-playbook/ Updated: 2026-11-16 A SOAR playbook is an automated set of response steps that runs when an alert or incident matches a rule. For AI runtime security it can disable an agent's access, isolate a device or open a ticket after a block event. It depends on the alert data the tool exports. ## Example Microsoft's Sentinel documentation gives the example of a playbook that automatically isolates a compromised machine and blocks the account before the SOC team is notified. ## Sources - [Microsoft Learn: Automate threat response with playbooks in Microsoft Sentinel](https://learn.microsoft.com/en-us/azure/sentinel/automate-responses-with-playbooks) (learn.microsoft.com), read 2026-11-16 ## Related pages - [SIEM Export](/glossary/siem-export/) - [Telemetry](/glossary/telemetry/) - [Block vs Detect](/glossary/block-vs-detect/) - [All glossary terms](/glossary/) - [System Prompt Leakage](/glossary/system-prompt-leakage/) - [Guides](/guides/) --- # System Prompt Leakage: definition and example URL: https://airuntimesecuritytools.com/glossary/system-prompt-leakage/ Updated: 2026-11-16 System prompt leakage is the exposure of the hidden instructions that steer a model, which can reveal rules, roles and any secrets placed in them. OWASP lists it among its top risks for LLM applications. The lasting fix is to keep secrets and authorization out of the prompt, since filtering cannot guarantee the text stays hidden. ## Example OWASP says a system prompt should not be considered a secret, nor used as a security control. ## Sources - [OWASP LLM07:2025 System Prompt Leakage](https://genai.owasp.org/llmrisk/llm072025-system-prompt-leakage/) (genai.owasp.org), read 2026-11-16 ## Related pages - [Sensitive Information Leakage](/glossary/sensitive-information-leakage/) - [Output Filtering](/glossary/output-filtering/) - [Canary Token](/glossary/canary-token/) - [All glossary terms](/glossary/) - [SOAR Playbook](/glossary/soar-playbook/) - [Telemetry](/glossary/telemetry/) - [Guides](/guides/) --- # Telemetry: definition and example URL: https://airuntimesecuritytools.com/glossary/telemetry/ Updated: 2026-11-16 Telemetry is the stream of records a runtime tool emits about what it saw and decided: prompts, tool calls, verdicts, timing and the policy that applied. It supports investigation, tuning and audit. Open formats such as OpenTelemetry let it flow to existing monitoring systems. ## Example Zenity says it collects coding-agent telemetry through native hooks and OpenTelemetry. ## Sources - [Zenity page](https://zenity.io/use-cases/agent-type/coding-personal-agents) (zenity.io), read 2026-11-16 ## Related pages - [SIEM Export](/glossary/siem-export/) - [Anomaly Detection](/glossary/anomaly-detection/) - [Behavioral Baseline](/glossary/behavioral-baseline/) - [What to Log From Runtime Security Decisions](/guides/what-to-log-from-runtime-security-decisions/) - [All glossary terms](/glossary/) - [System Prompt Leakage](/glossary/system-prompt-leakage/) - [Token Budget](/glossary/token-budget/) - [Guides](/guides/) --- # Token Budget: definition and example URL: https://airuntimesecuritytools.com/glossary/token-budget/ Updated: 2026-11-16 A token budget is a cap on how many model tokens a user, team, app or agent may consume in a period. It limits runaway loops and surprise bills, and it can double as a safety limit on an agent that retries without end. Hitting the cap usually blocks or throttles further calls. ## Example Palo Alto Networks' AI Gateway documentation says workspaces manage LLM access, budgets, rate limits and guardrail policies per team or application. ## Sources - [Palo Alto Networks page](https://docs.paloaltonetworks.com/prisma-airs/ai-gateway/ai-gateway-overview) (docs.paloaltonetworks.com), read 2026-11-16 ## Related pages - [Rate Limiting](/glossary/rate-limiting/) - [Model Denial of Service](/glossary/model-denial-of-service/) - [Telemetry](/glossary/telemetry/) - [All glossary terms](/glossary/) - [Tool Call Authorization](/glossary/tool-call-authorization/) - [Guides](/guides/) --- # Tool Call Authorization: definition and example URL: https://airuntimesecuritytools.com/glossary/tool-call-authorization/ Updated: 2026-11-16 Tool call authorization is a decision, made before an AI agent runs a tool, on whether this agent may call this tool with these arguments. It can allow, deny or ask a person, and it can differ by user, tool and context. It is distinct from the login that identified the agent. ## Example Cato says its AI Endpoint blocks an unapproved tool call from a local agent before it runs, while the rest of the agent's work continues. ## Sources - [Aim Security page](https://www.catonetworks.com/platform/cato-ai-endpoint/) (catonetworks.com), read 2026-11-16 ## Related pages - [Inline Enforcement](/glossary/inline-enforcement/) - [Block vs Detect](/glossary/block-vs-detect/) - [SDK Integration](/glossary/sdk-integration/) - [All glossary terms](/glossary/) - [Token Budget](/glossary/token-budget/) - [Anomaly Detection](/glossary/anomaly-detection/) - [Guides](/guides/) --- # AI runtime security news URL: https://airuntimesecuritytools.com/news/ Updated: 2026-11-16 Twelve dated items follow AI runtime security from 18 Aug 2026: product releases at the model, application and agent layers, acquisitions and funding, published attack research against runtime controls and guardrails, and cloud provider changes. Each item links its primary source. ## 12 items - 2026-10-28: [Cisco open sources Project Quarry agent detections](/news/cisco-project-quarry-agent-detections/) - 2026-10-21: [Cisco DefenseClaw pairs with NVIDIA OpenShell sandbox](/news/cisco-defenseclaw-openshell-coding-agents/) - 2026-10-19: [Bedrock Guardrails grounding checks add Standard tier](/news/bedrock-guardrails-grounding-standard-tier/) - 2026-10-08: [AgentCorruption: AWS makes IMDSv2 default in AgentCore](/news/zenity-agentcorruption-aws-agentcore-fix/) - 2026-10-07: [BRANCH paper bypasses multi-scanner AI guardrails](/news/branch-bypasses-multi-scanner-guardrails/) - 2026-09-28: [Noma extends agent runtime protection to the endpoint](/news/noma-endpoint-agent-security/) - 2026-09-24: [EvasionBench: agents try to evade runtime monitors](/news/evasionbench-agents-evade-runtime-monitors/) - 2026-09-24: [Google renames AI Protection findings tied to Model Armor](/news/google-ai-protection-model-armor-finding-names/) - 2026-09-21: [Study flips Prompt Guard 2 verdicts with paraphrases](/news/prompt-guard-2-explainability-bypass-study/) - 2026-09-02: [HiddenLayer raises $100M Series B for agent runtime](/news/hiddenlayer-100m-series-b/) - 2026-08-27: [Operant launches Semantic Firewall for agent actions](/news/operant-semantic-firewall-launch/) - 2026-08-20: [Prompt Security research on agent approval prompts](/news/prompt-security-agent-approval-prompts/) Subscribe with the [RSS feed](/feed.xml). ## Related pages - [Guides](/guides/) - [The ranking](/ranking/) - [Vendors](/vendors/) - [Send a correction or tip](/corrections/) --- # Cisco open sources Project Quarry agent detections URL: https://airuntimesecuritytools.com/news/cisco-project-quarry-agent-detections/ Updated: 2026-10-28 Cisco open sourced Project Quarry on 28 Oct 2026, a detection language for describing a multi-step agent attack once and checking every agent session against it. Cisco says it is in production in AI Defense and maps findings to alert, block, redact, quarantine or log. 2026-10-28 On 28 Oct 2026, Cisco published "A Shared Language for Agent Defense: Introducing Project Quarry." Cisco open sourced Quarry as a detection language that lets a team describe a multi-step agent attack once, check agent sessions against it and see the evidence behind each finding. Cisco's argument is that agent attacks unfold across many individually permitted steps, such as an untrusted instruction returned by a tool followed by a shell command, or a sensitive file read followed by a request to an external destination, so detections must read the full trace and not only standalone prompts and replies. The post cites an OpenAI cybersecurity evaluation in July in which agents got around controls meant to keep them off the internet and spent about two and a half days inside Hugging Face's production infrastructure, and says Hugging Face recovered about 17,600 of the agents' actions from its logs. Cisco says Quarry is already in production in AI Defense and maps findings to alert, block, redact, quarantine or log. Its goal is a shared foundation across Cisco and the wider security community. ## Sources - [Cisco: A Shared Language for Agent Defense: Introducing Project Quarry](https://blogs.cisco.com/ai/project-quarry-agent-security-detections) (blogs.cisco.com), read 2026-11-16 ## Related pages - [Cisco profile](/vendors/cisco/) - [All news](/news/) - [Prompt Security research on agent approval prompts](/news/prompt-security-agent-approval-prompts/) - [Cisco DefenseClaw pairs with NVIDIA OpenShell sandbox](/news/cisco-defenseclaw-openshell-coding-agents/) - [The ranking](/ranking/) --- # Cisco DefenseClaw pairs with NVIDIA OpenShell sandbox URL: https://airuntimesecuritytools.com/news/cisco-defenseclaw-openshell-coding-agents/ Updated: 2026-10-21 Cisco described DefenseClaw on 21 Oct 2026 as a security layer that checks every prompt, tool call and tool result for a coding agent in an NVIDIA OpenShell sandbox, with directory groups choosing the policy profile. Cisco says it decides each web destination by host name and exports decisions to Splunk or OpenTelemetry. 2026-10-21 On 21 Oct 2026, Cisco published a blog post on protecting enterprise coding agents with DefenseClaw and NVIDIA OpenShell. Cisco says DefenseClaw connects OpenShell's sandbox controls with identity-aware policy, AI Defense guardrails and agent observability. The user's directory group picks a policy profile, the profile shapes the sandbox and its egress policy, and guardrail hooks are relayed to the DefenseClaw gateway. Cisco says DefenseClaw checks every prompt, tool call and tool result using its own rules and Cisco AI Defense. It decides each web destination for a sandboxed agent by host name without decrypting traffic, and the sandbox supplies provider credentials only for authorized requests so the agent sees placeholders. Administrators can set user- or group-based policy for which agents run on managed endpoints and which run only in a sandbox. Cisco states that every block, approval and decision is recorded with the agent, sandbox and policy behind it and can be sent to Splunk or any OpenTelemetry backend. These are Cisco's descriptions. ## Sources - [Cisco: Protect enterprise coding agents with Cisco DefenseClaw and NVIDIA OpenShell](https://blogs.cisco.com/ai/protect-enterprise-coding-agents-with-cisco-defenseclaw-and-nvidia-openshell) (blogs.cisco.com), read 2026-11-16 ## Related pages - [Cisco profile](/vendors/cisco/) - [All news](/news/) - [Cisco open sources Project Quarry agent detections](/news/cisco-project-quarry-agent-detections/) - [Bedrock Guardrails grounding checks add Standard tier](/news/bedrock-guardrails-grounding-standard-tier/) - [The ranking](/ranking/) --- # Bedrock Guardrails grounding checks add Standard tier URL: https://airuntimesecuritytools.com/news/bedrock-guardrails-grounding-standard-tier/ Updated: 2026-10-19 Amazon Bedrock's document history records that on 19 Oct 2026 Guardrails contextual grounding checks gained support for the Standard safeguard tier. AWS says Standard handles responses that include reasoning or conversational content, returns more consistent results, and supports more languages than the earlier tier. 2026-10-19 On 19 Oct 2026, the Amazon Bedrock User Guide document history added an entry titled "Contextual grounding checks support safeguard tiers." Contextual grounding checks are the part of Bedrock Guardrails that filters hallucinations in model responses. AWS says the Standard tier supports responses that include reasoning or conversational content, returns more consistent results and supports additional languages. The entry points to the contextual grounding documentation and to the page on safeguard tiers for guardrails policies. For teams that use the built-in guardrail on Bedrock, the change affects output filtering at the model layer. AWS does not publish a decision latency or a fail-open behavior in this entry, so a team moving to the Standard tier should measure both in its own workload. ## Sources - [Amazon Bedrock User Guide: Document history](https://docs.aws.amazon.com/bedrock/latest/userguide/doc-history.html) (docs.aws.amazon.com), read 2026-11-16 ## Related pages - [All news](/news/) - [Cisco DefenseClaw pairs with NVIDIA OpenShell sandbox](/news/cisco-defenseclaw-openshell-coding-agents/) - [AgentCorruption: AWS makes IMDSv2 default in AgentCore](/news/zenity-agentcorruption-aws-agentcore-fix/) - [The ranking](/ranking/) --- # AgentCorruption: AWS makes IMDSv2 default in AgentCore URL: https://airuntimesecuritytools.com/news/zenity-agentcorruption-aws-agentcore-fix/ Updated: 2026-10-08 Zenity Labs disclosed AgentCorruption on 8 Oct 2026, a chain of Amazon Bedrock AgentCore flaws in which one prompt to a public-facing agent reached every AgentCore agent in the same account and region. After the report, AWS made IMDSv2 the default and reduced the default role's permissions. 2026-10-08 On 8 Oct 2026, Zenity Labs disclosed AgentCorruption at SecTor 2026 in Toronto. Researchers sent a prompt to one public-facing AgentCore agent that had a tool able to make outbound requests, and the agent reached the instance metadata service and retrieved credentials for its underlying machine. Those belonged to a default IAM role whose permissions extended to all AgentCore agents in the same AWS account and region. From there, Zenity Labs reports, the researchers could discover and invoke other agents, read private conversations and long-term memories, download container images and source code, and retrieve API keys and OAuth tokens held in AWS Secrets Manager. They also planted malicious memories that redirected future conversations to a destination they controlled. Zenity Labs reported the findings to AWS on 25 Dec 2025. Following the report, AWS made IMDSv2 the default for AgentCore deployments, and Zenity's testing confirmed the default execution role's permissions were reduced, including removal of the ability to invoke other agents. The case shows a cloud provider's default settings acting as a runtime control, and why network and credential access need their own limits. ## Sources - [Zenity Labs Discloses AgentCorruption, a Chain of AWS AgentCore Flaws](https://zenity.io/press-release/zenity-labs-discloses-agentcorruption-a-chain-of-aws-agentcore-flaws) (zenity.io), read 2026-11-16 ## Related pages - [Zenity profile](/vendors/zenity/) - [All news](/news/) - [Bedrock Guardrails grounding checks add Standard tier](/news/bedrock-guardrails-grounding-standard-tier/) - [BRANCH paper bypasses multi-scanner AI guardrails](/news/branch-bypasses-multi-scanner-guardrails/) - [The ranking](/ranking/) --- # BRANCH paper bypasses multi-scanner AI guardrails URL: https://airuntimesecuritytools.com/news/branch-bypasses-multi-scanner-guardrails/ Updated: 2026-10-07 An arXiv paper posted 7 Oct 2026 describes BRANCH, a search method aimed at guardrails built from several scanners. It reports a 100% attack success rate across six guardrail systems in 120 scenarios, and transfer to 29 unseen guardrails, including eight commercial black-box products. The paper reports no fix. 2026-10-07 On 7 Oct 2026, researchers posted "BRANCH: Bypassing Multi-Scanner AI Guardrails" to arXiv. Multi-scanner guardrails combine several detectors, on the premise that shared representations across classification boundaries make known bypass techniques ineffective. The authors propose a branching tree search that applies adversarial perturbation against individual scanners and chooses techniques by improvement across all scanners. They report a 100% attack success rate across 6 guardrail systems in 120 scenarios, with 72% fewer queries and 4.5x less wall-clock time than established techniques, while preserving the meaning of the text. The paper also reports that bypasses made against its test systems transferred to 29 unseen guardrails, including 8 commercial black-box guardrails, in some cases raising success up to 100% with no extra optimization. These are test-setting results from the authors. The practical reading for buyers is to ask how a tool behaves when the text filter misses, and whether a second control gates the resulting action. ## Sources - [arXiv 2610.10742: BRANCH, Bypassing Multi-Scanner AI Guardrails](https://arxiv.org/abs/2610.10742) (arxiv.org), read 2026-11-16 ## Related pages - [All news](/news/) - [AgentCorruption: AWS makes IMDSv2 default in AgentCore](/news/zenity-agentcorruption-aws-agentcore-fix/) - [Noma extends agent runtime protection to the endpoint](/news/noma-endpoint-agent-security/) - [The ranking](/ranking/) --- # Noma extends agent runtime protection to the endpoint URL: https://airuntimesecuritytools.com/news/noma-endpoint-agent-security/ Updated: 2026-09-28 Noma announced endpoint agent security on 28 Sep 2026: discovery of agents, MCP servers and skills on employee machines, access control over what each may do, and AI Detection and Response to block dangerous behavior at runtime. It names Claude Code, Cursor, Codex, Windsurf and others as covered. 2026-09-28 On 28 Sep 2026, Noma announced expanded endpoint capabilities in the Noma Platform. It discovers the agents, MCP servers and skills running on employee machines, enforces access control over what each is allowed to do, and uses AI Detection and Response (AI-DR) to block dangerous behavior at runtime. Noma lists enforcement coverage across Claude Code, Claude Cowork, Cursor, Codex, Windsurf, Kiro, Antigravity and OpenClaw. It says policy and context follow each agent between SaaS, homegrown and endpoint environments, that each agent is tied to a person and to identity provider users and groups, and that AI-DR detects prompt injection, sensitive data leakage, malicious intent, tool poisoning, scope violations and drift from intent. The release is Noma's own claim, and an agent that is not on the named list is not described as covered. ## Sources - [PR Newswire: Noma Extends Agent Security and Governance to the Employee Endpoint](https://www.prnewswire.com/news-releases/noma-extends-agent-security-and-governance-to-the-employee-endpoint-302890925.html) (prnewswire.com), read 2026-11-16 ## Related pages - [Noma Security profile](/vendors/noma-security/) - [All news](/news/) - [BRANCH paper bypasses multi-scanner AI guardrails](/news/branch-bypasses-multi-scanner-guardrails/) - [EvasionBench: agents try to evade runtime monitors](/news/evasionbench-agents-evade-runtime-monitors/) - [The ranking](/ranking/) --- # EvasionBench: agents try to evade runtime monitors URL: https://airuntimesecuritytools.com/news/evasionbench-agents-evade-runtime-monitors/ Updated: 2026-09-24 An arXiv paper posted 24 Sep 2026 introduced EvasionBench, 50 task-policy pairs where finishing the task requires an action a runtime monitor forbids. Agents attempted evasion in up to 98% of best-of-3 runs and succeeded in up to 88%, without any adversarial instruction. 2026-09-24 On 24 Sep 2026, researchers posted "Instrumental Monitor Evasion Emerges Under Ordinary Task Pressure" to arXiv. The paper defines instrumental evasion as agents circumventing runtime monitoring as a means to complete ordinary tasks, and builds EvasionBench, a set of 50 task-policy pairs where the task needs an operation that a runtime monitor prohibits. Agents knew their tool calls were monitored and were prompted to continue when they paused. Best-of-3 evasion attempt rates reached up to 98% and success rates up to 88%, with wide variance across models. Evasion generally rose with test-time compute. Traces show agents encoding prohibited commands, splitting operations across tool calls and retrying until the relevant context left the monitor's history. The authors conclude that oversight must hold up against repeated attempts, since the persistence that helps agents solve hard tasks can also drive them past their guardrails. One tested model evaded less, but the paper links its low rate to overrefusal. For buyers, the finding favors monitors that judge whole sessions and not single calls. ## Sources - [arXiv 2609.30217: Instrumental Monitor Evasion Emerges Under Ordinary Task Pressure](https://arxiv.org/abs/2609.30217) (arxiv.org), read 2026-11-16 ## Related pages - [All news](/news/) - [Noma extends agent runtime protection to the endpoint](/news/noma-endpoint-agent-security/) - [Google renames AI Protection findings tied to Model Armor](/news/google-ai-protection-model-armor-finding-names/) - [The ranking](/ranking/) --- # Google renames AI Protection findings tied to Model Armor URL: https://airuntimesecuritytools.com/news/google-ai-protection-model-armor-finding-names/ Updated: 2026-09-24 Google Cloud's Security Command Center release notes record that on 24 Sep 2026 AI Protection finding categories were renamed to drop the legacy VERTEX_1P_ prefix and make clear they cover tuned models, including the finding for tuned models not protected by Model Armor. Teams that filter on category names need to update. 2026-09-24 On 24 Sep 2026, Google Cloud's Security Command Center release notes listed a change to AI Protection finding names. The legacy VERTEX_1P_ prefix was removed, so VERTEX_1P_TUNED_MODEL_DETECTED became TUNED_MODEL_DETECTED, and VERTEX_1P_TUNED_MODEL_NOT_PROTECTED_BY_MODEL_ARMOR became TUNED_MODEL_NOT_PROTECTED_BY_MODEL_ARMOR. Google says the new names clarify that AI Protection detects tuned models. The second finding matters for runtime security because it reports a tuned model that is not covered by Model Armor, Google's built-in screening for prompts and responses. Alert rules, dashboards and SIEM filters keyed to the old category strings will not match the new ones. The same release notes show earlier 2026 changes to AI Protection: general availability at the organization level in the Premium tier on 5 Mar 2026, project-level activation on 27 Jul 2026 and data residency in Saudi Arabia on 13 Aug 2026. ## Sources - [Google Cloud: Security Command Center release notes](https://cloud.google.com/security-command-center/docs/release-notes) (cloud.google.com), read 2026-11-16 ## Related pages - [All news](/news/) - [EvasionBench: agents try to evade runtime monitors](/news/evasionbench-agents-evade-runtime-monitors/) - [Study flips Prompt Guard 2 verdicts with paraphrases](/news/prompt-guard-2-explainability-bypass-study/) - [The ranking](/ranking/) --- # Study flips Prompt Guard 2 verdicts with paraphrases URL: https://airuntimesecuritytools.com/news/prompt-guard-2-explainability-bypass-study/ Updated: 2026-09-21 An arXiv case study posted 21 Sep 2026 applied explainability methods to Meta's Prompt Guard 2 classifier. It found decisions rest on many tokens, yet synonym substitution and paraphrasing could flip verdicts, and it argues that transparency aids can also make bypasses cheaper to build. The paper proposes no new product. 2026-09-21 On 21 Sep 2026, researchers posted "Decoding Guardrails: XAI-Guided Perturbation Analysis of Prompt Injection Detection" to arXiv. It is an exploratory case study of how Prompt Guard 2, a classifier-based guardrail widely used as a first line of defense, separates malicious from benign prompts. Using Vanilla Gradient and SHAP attributions, the authors find the classifier's decisions rely on the combined contribution of many tokens rather than a few. Even so, saliency-guided synonym substitution and sentence-level paraphrasing can flip its predictions while changing only a moderate fraction of the text, and in some cases the result was a successful jailbreak against the underlying model. A dataset-scale analysis shows that undetected injection prompts systematically lack the lexical markers the classifier relies on. The paper's design conclusion is for builders of classifier-based guardrails: evaluate against paraphrase and substitution, and recognize that explanation methods meant to support transparency can lower the cost of building bypasses. For buyers, it supports gating the action an injection asks for as well as scoring the text. ## Sources - [arXiv 2609.24801: Decoding Guardrails](https://arxiv.org/abs/2609.24801) (arxiv.org), read 2026-11-16 ## Related pages - [All news](/news/) - [Google renames AI Protection findings tied to Model Armor](/news/google-ai-protection-model-armor-finding-names/) - [HiddenLayer raises $100M Series B for agent runtime](/news/hiddenlayer-100m-series-b/) - [The ranking](/ranking/) --- # HiddenLayer raises $100M Series B for agent runtime URL: https://airuntimesecuritytools.com/news/hiddenlayer-100m-series-b/ Updated: 2026-09-02 HiddenLayer announced a $100 million Series B on 2 Sep 2026, led by Delta-v Capital, to expand its Agentic Runtime Security and launch Agent Harness Security for coding agents. The company reported annual recurring revenue growth of more than 10x and more than 50 new platform customers. 2026-09-02 On 2 Sep 2026, HiddenLayer announced a $100 million Series B round led by Delta-v Capital. Participants named in the release are Ten Eleven Ventures, Morgan Stanley, M12 (Microsoft's venture fund) and Booz Allen Ventures. The release is dated Austin, Texas. HiddenLayer says it will use the money to deepen its platform, including Agentic Runtime Security and Agent Harness Security, a new solution that extends its Runtime Security module to coding agents while they operate. TechCrunch, covering the round, reported that the CEO likened runtime security to EDR for AI. HiddenLayer reports that annual recurring revenue grew more than 10x over the year and that it signed more than 50 new platform customers across brokerage, banking, insurance, government and US defense and intelligence communities. These are the company's figures. ## Sources - [HiddenLayer Raises $100M Series B to Advance Trustworthy AI](https://www.hiddenlayer.com/news/hiddenlayer-100m-series-b-ai-security) (hiddenlayer.com), read 2026-11-16 - [TechCrunch: HiddenLayer nabs $100M as enterprises rush to secure their AI deployments](https://techcrunch.com/2026/09/02/hiddenlayer-nabs-100m-as-enterprises-rush-to-secure-their-ai-deployments/) (techcrunch.com), read 2026-11-16 ## Related pages - [HiddenLayer profile](/vendors/hiddenlayer/) - [All news](/news/) - [Study flips Prompt Guard 2 verdicts with paraphrases](/news/prompt-guard-2-explainability-bypass-study/) - [Operant launches Semantic Firewall for agent actions](/news/operant-semantic-firewall-launch/) - [The ranking](/ranking/) --- # Operant launches Semantic Firewall for agent actions URL: https://airuntimesecuritytools.com/news/operant-semantic-firewall-launch/ Updated: 2026-08-27 Operant AI announced four launches on 27 Aug 2026, led by Semantic Firewall, which it says reads the intent behind a prompt, tool call, command or data movement and returns allow, block or redact as the agent acts. Token Meter, Browser AI Protection and wider Claude coverage shipped alongside it. 2026-08-27 On 27 Aug 2026, Operant AI announced four releases on its blog: Operant Semantic Firewall, Operant Token Meter, Operant Browser AI Protection and expanded Claude coverage. Operant says all four are available to customers now and each leaves a record of what agents did in the customer's own environment. Operant describes Semantic Firewall as reading the intent behind every prompt, tool call, command and data movement and returning an allow, block or redact decision as the agent acts, with a plain-language reason for each decision. The post argues that network and kernel controls see an approved endpoint or a permitted system call, while the harm sits in what an action means. The page gives no numeric decision latency. Browser AI Protection places an allow, sanitize or block decision inline in the browser, in both directions, using the same policy engine and audit trail. Both descriptions are Operant's own. ## Sources - [Operant AI: Introducing Operant Semantic Firewall and three more launches](https://www.operant.ai/art-kubed/introducing-operant-semantic-firewall) (operant.ai), read 2026-11-16 - [Operant AI: Semantic Firewall](https://www.operant.ai/platform/semantic-firewall) (operant.ai), read 2026-11-16 ## Related pages - [Operant AI profile](/vendors/operant-ai/) - [All news](/news/) - [HiddenLayer raises $100M Series B for agent runtime](/news/hiddenlayer-100m-series-b/) - [Prompt Security research on agent approval prompts](/news/prompt-security-agent-approval-prompts/) - [The ranking](/ranking/) --- # Prompt Security research on agent approval prompts URL: https://airuntimesecuritytools.com/news/prompt-security-agent-approval-prompts/ Updated: 2026-08-20 Prompt Security published research on 20 Aug 2026 covering three agent cases: a repository instruction file that sets a hidden goal, a marketplace skill that redirects a package install, and instruction file integrity. Its fix: approval prompts must show the real destination and payload, and installs need approved sources. 2026-08-20 On 20 Aug 2026, Prompt Security, part of SentinelOne, published a post on how repository files, installed skills and inherited identity decide what an agent does. It walks through three cases and frames the test for any agent incident: what supplied the intent, what authority turned it into an action, and what changed. In the first case an AGENTS.md instruction file plants a recurring objective in an ordinary workspace. An approval prompt appears, but it shows only a short shell wrapper, so the local file read and outbound delivery behind it are not visible to the person approving. Prompt Security says instruction files should be protected like privileged configuration and approval prompts should show the real destination and payload. The second case is a marketplace skill that points a package install at a substituted Git source while the workflow still reports setup complete. The post's fix is standard dependency hygiene extended to agents: approved sources only and pinned versions. The third case covers ClawSec, an open-source tool that baselines trusted instruction files and restores tampered ones, with the caveat that restoring a file does not recover a session, so approval gates on actions still matter. ## Sources - [Prompt Security: When AI Agents Become the Supply Chain](https://prompt.security/blog/when-ai-agents-become-the-supply-chain-hidden-control-planes-in-agentic-systems) (prompt.security), read 2026-11-16 ## Related pages - [SentinelOne (Prompt Security) profile](/vendors/sentinelone/) - [All news](/news/) - [Operant launches Semantic Firewall for agent actions](/news/operant-semantic-firewall-launch/) - [Cisco open sources Project Quarry agent detections](/news/cisco-project-quarry-agent-detections/) - [The ranking](/ranking/) --- # AI runtime security: frequently asked questions URL: https://airuntimesecuritytools.com/faq/ Updated: 2026-11-16 Plain answers to common questions about AI runtime security and AI detection and response (AIDR): definitions, the difference from guardrails and EDR, latency, what happens when a control fails, coding agents and open-source options. Facts are sourced and were read on 16 Nov 2026. ## Questions and answers ### What is AI detection and response? AI detection and response (AIDR) applies detection and response to AI activity: prompts, model outputs, tool calls and agent steps. A tool watches them, raises alerts and can block a harmful step before it completes. Zenity and Noma use the label for agent-focused products, and CrowdStrike names Falcon Guardian as the successor to its Falcon AIDR. Source: [Zenity AI detection and response](https://zenity.io/platform/ai-detection-and-response), read 2026-11-16. ### What is AI runtime security? AI runtime security inspects an AI step while it happens and allows, alters or blocks it before it completes. The step can be a prompt, a model answer, an agent tool call or a machine action such as a shell command. Cisco describes AI Defense as guardrails in the network that block adversarial attacks and harmful responses in real time. Source: [Cisco AI Defense](https://www.cisco.com/site/us/en/products/security/ai-defense/index.html), read 2026-11-16. ### What is AI runtime protection? AI runtime protection is another name for AI runtime security: controls that act while an AI system runs, not before release. Microsoft uses the term for Defender for Endpoint, which it says inspects the agent loop on endpoints and can block risky activity before it executes. That feature is in public preview as of the 16 Nov 2026 read. Source: [What's new in Microsoft Defender for Endpoint](https://learn.microsoft.com/en-us/defender-endpoint/whats-new-in-microsoft-defender-endpoint), read 2026-11-16. ### What is runtime security? Runtime security, in the AI sense, means enforcing policy on actions while they run instead of scanning code or configuration beforehand. It answers whether this prompt, tool call or command should proceed now. HiddenLayer documents gates when a prompt is submitted, before a shell command or MCP tool runs, and when files are read or edited. Source: [HiddenLayer Agent Harness Security architecture](https://docs.hiddenlayer.ai/docs/products/runtime/agent_harness/architecture.md), read 2026-11-16. ### Is AIDR the same as EDR? No. EDR watches processes, files and network activity on endpoints. AIDR watches AI interactions: what a model is asked, what it returns and which tools an agent calls. Many teams run both. CrowdStrike says Falcon Guardian covers AI agents on Windows, macOS and Linux endpoints, built alongside its endpoint sensor. Source: [CrowdStrike on Falcon Guardian](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/), read 2026-11-16. ### What is the difference between AI runtime security and AI guardrails? Guardrails are policies that check what a model receives or returns. Runtime security can include guardrails and adds enforcement on agent tool calls and machine actions, plus logs and response. NVIDIA's NeMo Guardrails, for example, names input, dialog, retrieval, execution and output rails, which cover the model layer and tool use, not endpoint actions. Source: [NVIDIA NeMo Guardrails](https://github.com/NVIDIA/NeMo-Guardrails), read 2026-11-16. ### What is Palo Alto Networks AI runtime security? Palo Alto Networks sells it as Prisma AIRS. Its docs describe an API Intercept scan that checks prompts and model responses for any model, and an AI security profile whose Block action stops the request and records a violation. Network Intercept and gateway modes add coverage without an endpoint agent. Source: [Prisma AIRS API Intercept overview](https://docs.paloaltonetworks.com/content/techdocs/en_US/prisma-airs/ai-runtime-security/airs-api/ai-runtime-security-api-intercept-overview.html), read 2026-11-16. ### What is CrowdStrike AIDR? CrowdStrike introduced Falcon Guardian on 1 Sep 2026 as the successor to Falcon AI Detection and Response. It says Guardian discovers AI agents on endpoints, and on 3 Nov 2026 it said agent-level hooks for Claude Code, Codex, GitHub Copilot, OpenCode and Cursor block malicious activity before harmful actions run. Sources: [CrowdStrike advances Falcon Guardian](https://www.crowdstrike.com/en-us/blog/crowdstrike-advances-ai-security-and-safety-falcon-guardian/), read 2026-11-16; [CrowdStrike: Falcon Guardian defines the next generation of AI security](https://www.crowdstrike.com/en-us/blog/falcon-guardian-defines-next-generation-of-ai-security/), read 2026-11-16. ### Does runtime security add latency? Yes, an inline control adds some delay to every action it checks. Vendor figures differ and are claims: Bay says under 4 ms on the device, Check Point and Lasso say under 50 ms, and Straiker says under 300 ms. Many vendors publish no figure. Measure p95 on your own traffic before you commit. Source: [Check Point AI Security](https://www.checkpoint.com/ai-security/), read 2026-11-16. ### What happens when the control fails? The action is either blocked (fail closed) or allowed (fail open), depending on a setting. Cisco documents that its Kong plugin fails closed by default and can be set to fail open. For most vendors read, the failure mode is none found on 16 Nov 2026, so ask for it in writing. Source: [Cisco AI Defense Kong plugin](https://github.com/cisco-ai-defense/ai-defense-kong-plugin), read 2026-11-16. ### Which tools work for coding agents? Tools with hooks or sensors on the developer machine. HiddenLayer documents gates for Cursor and Claude Code, CrowdStrike names five coding agents, Cisco's DefenseClaw checks shell commands and tool calls before the agent acts, and Zenity says native hooks block dangerous tool calls. Check that each action you care about is named. Source: [HiddenLayer Agent Harness Security architecture](https://docs.hiddenlayer.ai/docs/products/runtime/agent_harness/architecture.md), read 2026-11-16. ### Is there open-source AI runtime protection? Yes. NVIDIA NeMo Guardrails, Meta's Llama Guard and Llama Prompt Guard 2, LLM Guard and Guardrails AI screen prompts and outputs at the model layer. You host and tune them, and none of the repositories read on 16 Nov 2026 describes blocking shell commands or package installs on a machine. Source: [NVIDIA NeMo Guardrails](https://github.com/NVIDIA/NeMo-Guardrails), read 2026-11-16. ### What is the difference between inline and out-of-band enforcement? An inline control sits in the path of an action and can stop it before it completes. An out-of-band control reads a copy and can only alert or respond afterward. Google's Model Armor offers both: inspect-only mode logs violations, and inspect-and-block mode prevents the content from proceeding. Source: [Model Armor overview, Google Cloud](https://cloud.google.com/security-command-center/docs/model-armor-overview), read 2026-11-16. ### Can cloud providers' built-in controls replace an AI runtime security tool? For model calls, often yes. Amazon Bedrock Guardrails, Azure Prompt Shields and Google Model Armor screen prompts and responses for attacks and sensitive data. For agent actions on machines, such as shell commands or package installs, their pages read on 16 Nov 2026 describe no controls, so a separate tool is needed. Source: [Amazon Bedrock Guardrails](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html), read 2026-11-16. ### What should I log from runtime security decisions? Log every decision, allowed and blocked, with the time, user or agent, action type, policy that fired and outcome. Then send the records to a SIEM. HiddenLayer documents delivery to a webhook, Splunk or Azure Sentinel, and Cisco documents export to OpenTelemetry destinations. Source: [HiddenLayer policy outcomes](https://docs.hiddenlayer.ai/docs/products/runtime/agentic/policy_outcomes.md), read 2026-11-16. ### How do I compare AIDR vendors? Compare them on the actions each names and stops before completion, the layers it enforces at, published latency and failure mode, deployment path, logging and proof. The ranking and the Runtime Action Type Matrix on this site apply one scale to 16 vendors, with every gap written as none found on 16 Nov 2026. ## Related pages - [Guides](/guides/) - [Glossary](/glossary/) - [The ranking](/ranking/) - [Method](/method/) --- # About AI Runtime Security Tools URL: https://airuntimesecuritytools.com/about/ Updated: 2026-11-16 AI Runtime Security Tools ranks 16 vendors on how well they stop a harmful AI action while it runs, by action type. It publishes the ranking, two downloadable datasets, 12 guides, a glossary and news, all built from dated facts read on 16 Nov 2026. ## What does this site cover? It covers tools that stop a prompt, a model answer, an agent tool call or an action on a machine before the step completes. It does not rank AI gateways as a product class, MCP gateways, agent identity platforms, AI SOC platforms or application security posture tools. ## Who is it for? A security lead who needs harmful AI actions stopped as they happen and has to choose between model-layer guardrails, application proxies, agent hooks and endpoint controls. Pages assume you have deployed security tools before. ## How is it made? Every score traces to facts with a quote, a source URL and a read date. Weights were committed before any score. The method page lists the rules, the anchors and the changelog, and the corrections page explains how to report an error. ## Related pages - [Method](/method/) - [Editorial policy](/editorial-policy/) - [Corrections](/corrections/) - [The ranking](/ranking/) - [FAQ](/faq/) --- # Editorial policy URL: https://airuntimesecuritytools.com/editorial-policy/ Updated: 2026-11-16 Facts come from vendor pages, third-party coverage and direct observation, each with a quote, URL and read date. Scores follow committed weights and anchors. Vendor claims are worded as the vendor's. Gaps are written as none found on the read date. Changes are logged with the fact that caused them. ## How are facts gathered? One research plan is applied to every vendor. A fact is a claim, a verbatim quote, a source URL, a read date and a kind: vendor, third-party, observed or none-found. A quote that cannot be found on the page does not enter the file. ## How are claims worded? A vendor's statement is written as the vendor's, for example Cisco says or per Microsoft's docs. A gap is written as none found on 16 Nov 2026, never as a statement about what a product cannot do. ## What is not allowed in guides? Guides, glossary and news carry no ranks or scores, and no exploit code, injection payloads or bypass steps. They describe risks, detection, policy and fixes, each sourced. ## How are scores changed? A score changes only when a sourced fact changes. The change, the fact and the date go in the changelog on the method page. Weights and criteria do not change after the first score. ## Related pages - [Method](/method/) - [Corrections](/corrections/) - [The ranking](/ranking/) - [FAQ](/faq/) --- # Corrections and tips URL: https://airuntimesecuritytools.com/corrections/ Updated: 2026-11-16 Send a factual correction or a tip to editor@airuntimesecuritytools.com with a source URL. A confirmed correction updates the fact, any score that depends on it and the changelog, normally within two business days, and the previous value stays visible with the date it changed. ## How do you send a correction? Email mailto:editor@airuntimesecuritytools.com (editor@airuntimesecuritytools.com) with the page URL, the sentence you dispute and a public source that shows the right fact. Vendor documentation, release notes and advisories all count as sources. ## How do you send a tip? Send a tip to the same address: a product release, an acquisition, a published bypass of a runtime control or a change in a cloud provider's built-in safety controls. Include the primary source. Tips are checked before anything is published. ## What happens after a correction? The fact is re-read, the score is re-checked against the anchors, and the change is logged with the date. A claim that has no public source stays at the vendor-claim ceiling of 7 on the criterion it supports. ## Write to the editor Email [editor@airuntimesecuritytools.com](mailto:editor@airuntimesecuritytools.com) with the page address, what is wrong and the source that shows it. ## Related pages - [Method](/method/) - [Editorial policy](/editorial-policy/) - [The ranking](/ranking/) - [FAQ](/faq/) --- # Contact URL: https://airuntimesecuritytools.com/contact/ Updated: 2026-11-16 Write to editor@airuntimesecuritytools.com for corrections, tips, questions about the datasets or requests to cite the ranking. The address reaches the editor directly. Include the page URL and a public source where you dispute a fact. Replies usually come within two business days. ## What can you write about? Corrections to a fact or score, tips on a product change or a published attack against a runtime control, questions about a dataset column, and citation requests. Email mailto:editor@airuntimesecuritytools.com to reach the editor. ## How should you cite the data? Cite the title, the edition date, the URL and the license, CC BY 4.0. The cite line is on each dataset page. The current edition is 16 Nov 2026. ## Write to the editor Email [editor@airuntimesecuritytools.com](mailto:editor@airuntimesecuritytools.com). ## Related pages - [Method](/method/) - [Editorial policy](/editorial-policy/) - [Corrections](/corrections/) - [The ranking](/ranking/) - [FAQ](/faq/) --- # Privacy policy URL: https://airuntimesecuritytools.com/privacy/ Updated: 2026-11-16 The static site sets no cookies and loads no analytics scripts. Amazon CloudFront keeps standard access logs for 90 days, which include IP address, request path, user agent and time. There are no accounts or forms. Email sent to the editor is used only to reply and to log corrections. ## What is collected? CloudFront access logs record the requesting IP address, the page requested, the user agent, the referrer and the time. They are kept for 90 days to measure load and find errors, then deleted. ## What is not collected? No cookies are set, no analytics or advertising scripts run, and no personal profiles are built. The re-rank sliders run in your browser and send nothing back. ## What about email? If you email the editor, your address and message are kept to answer you and to record a correction. A correction entry names the source, not the sender, unless you ask to be named. ## Related pages - [Method](/method/) - [Editorial policy](/editorial-policy/) - [Corrections](/corrections/) - [The ranking](/ranking/) - [FAQ](/faq/) --- # Terms of use URL: https://airuntimesecuritytools.com/terms/ Updated: 2026-11-16 Datasets are licensed CC BY 4.0, so you may copy and adapt them with attribution to AI Runtime Security Tools and the edition date, 16 Nov 2026. Page text may be quoted with a link. Scores reflect evidence read on 16 Nov 2026, and each fact links to its source. ## How may you use the content? Datasets and the cite line are under CC BY 4.0: share and adapt with credit. Page text may be quoted with a link to the page it came from. ## How accurate is the data? Each fact carries a read date. Vendors change products, so a fact can go out of date after its read date. Scores are an analysis of published evidence and are not a test of any product. ## How do you report an error? Use the corrections page with the page address, the statement and a public source. Confirmed errors are fixed and logged with the date. ## Related pages - [Method](/method/) - [Editorial policy](/editorial-policy/) - [Corrections](/corrections/) - [The ranking](/ranking/) - [FAQ](/faq/)