Cisco AI runtime security
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.
Facts read on .
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.
| 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, cisco-14, cisco-15, cisco-49, cisco-51, cisco-02, cisco-20, 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, cisco-05, cisco-12, cisco-11, cisco-45 |
| Latency and failure mode | 5 of 10 | 15 | Timeouts and fail-closed defaults are published. Inline latency figure: none found. | cisco-06, cisco-07, cisco-09, 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, cisco-38, cisco-11, cisco-44, 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, cisco-18, cisco-35, cisco-53, 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, cisco-57, cisco-56 |
| Proof | 5 of 10 | 5 | A public company with one named customer voice. Analyst coverage: none found. | cisco-58, cisco-59, cisco-54 |
Where it ranks
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
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .
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, read .