Responding to an AI Agent Security Incident
A starting sequence for containing agent execution, preserving evidence, checking downstream impact, and recovering under controlled authority.
Contain the workflow. Reconstruct the action.
When an agent behaves unexpectedly, investigate the actions it actually took and the authority used for them. An unusual response may be the visible symptom of a tool call, memory change, or delegated task. Activate the organization's incident process and involve the owners of affected systems.
1. Stop further impact
Use the approved containment path to pause execution and disable affected tools. Check queued jobs, schedulers, child agents, and external automations. Revoke or restrict exposed credentials where appropriate. Record what was stopped, when it was stopped, and what remains active.
2. Preserve the evidence trail
Retain action and correlation IDs, tool arguments, authorization outcomes, relevant model and policy versions, and changes to memory or configuration. Protect evidence against alteration and limit access. Redact secrets in working copies; preserve sensitive originals only through the team's controlled evidence process.
3. Establish scope
Compare intended actions with actual writes, exports, privilege changes, and delegated operations. Ask which identities and downstream systems accepted the requests. Trace the first suspicious input and determine whether it affected later runs or shared state.
Example: a suspicious data export
Stopping an agent does not prove that an export was stopped. Check the destination service, queued transfer, credentials, and recipient. Preserve the transaction evidence and establish whether data left the approved boundary.
4. Repair the enforcement gap
Identify the boundary that allowed the unsafe action. Correct the relevant permission, tool behavior, approval flow, or state handling. Test the observed abuse path and nearby variations in an isolated environment before restoring access.
5. Restart with a recorded decision
Name the recovery owner, require a clean credential and state review, and verify that containment can be invoked again. Resume with limited permissions and increased observation. Capture follow-up actions and the date of the next review.
Prepare this process before deployment. NIST SP 800-61 Rev. 3 places incident response within ongoing cybersecurity risk management; agent-specific procedures should fit that broader process.