AI Agent Security Solutions

AI agent security incident response playbook

When an AI agent misbehaves, contain it first, revoke its credentials second, preserve its trail third, then find the cause and restore with tighter limits. Decide in advance who can pull the kill switch, which logs hold the agent's tool calls and how you cut tokens at the identity provider.

Facts read on .

What do you do when an agent misbehaves?

Stop the agent from acting, then work out what it did. Pause the session or cut its tool access before any analysis, because a running agent keeps acting.

Straiker says its kill switch cuts off tool access, suspends the session, isolates or takes an agent offline. CrowdStrike says its AI Kill Switch directs connected AI gateways. Check what your product can pause, and rehearse it.

Decide in advance who may pull the switch and under what conditions. A written rule such as any agent that reads credentials outside its scope is paused at once spares a debate at 2 a.m. Record the contact for each platform owner.

Sources: Straiker: agentic kill switch; CrowdStrike: AI Kill Switch.

How do you revoke an agent and review its trail?

Revoke at the identity provider and at each tool, then export the trail. Revoking only the agent leaves tokens that other services still honor.

Island describes Okta and Entra revocation steps for agent identities, and Zenity says it reads agent identities from Entra ID and Okta. Microsoft documents that Conditional Access for agents evaluates agent, user, device and risk signals to allow, block or require controls.

Revoke in the right order: stop the agent, revoke its tokens, rotate any secret it could read, then disable the agent identity. Check that services which cached the token also stop honoring it, since some keep a token valid until it expires.

Sources: Island: agentic identity; Zenity: agentic identity and access management; Microsoft Entra: Conditional Access for agents.

Where is the evidence?

The evidence sits in the agent platform's logs, the endpoint product's decision log, the gateway and the identity provider's sign-in logs. List them before an incident.

Bay describes session timelines for investigation, and Cisco says every block, approval and decision is recorded with the agent, sandbox and policy behind it. Export the records before retention rules remove them.

Preserve evidence before you change settings. Snapshot the agent configuration, the list of connected tools and the recent logs. If the platform keeps logs for a short time, export first and investigate second.

Sources: Bay: AI agent investigation; Cisco: protect enterprise coding agents with DefenseClaw.

How do you find the cause?

Trace the chain from the input to the action: what the agent read, what it decided, which tool it called and with what permission. The cause is usually one of OWASP's three: too much functionality, too much permission or too much autonomy.

Triggers include indirect prompt injection and, in multi-agent systems, a compromised peer agent.

Ask whether a person, a document, a web page or another agent supplied the input that changed the behavior. Where the input is external, add it to your blocked sources and test that the new rule fires.

Sources: OWASP LLM06:2025 Excessive Agency; NIST CAISI request for information on securing AI agent systems.

How do you restore safely?

Restore with narrower tools, narrower scopes and an approval step for the action that went wrong. Test the new limit with a harmless case.

Do not restore the old credential. Issue a new short-lived one.

Tell the users and owners of the agent what happened and what changed. If data left the company, follow your normal breach process; an agent incident is still a data incident.

Sources: RFC 9700, Best Current Practice for OAuth 2.0 Security.

What do you write down afterward?

Record the timeline, the control that should have stopped it, the control that did, the gap and the owner of the fix. Add the incident to the monthly report.

Use MITRE ATLAS technique names where they fit, so the record uses shared words.

Rehearse once a year with a tabletop. Use a real agent in your estate, a plausible failure and the actual contacts. Fix the gaps the exercise finds in the playbook itself.

Sources: MITRE ATLAS.

Sources