# 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/)
