AI incident response is what a team does when an AI system causes or nearly causes harm: contain it, work out what happened from the logs, fix the cause, and change policies so it does not recur.
What counts as an incident
- An agent deleted, sent or changed something it should not have.
- A response leaked personal data or another customer's information.
- A jailbreak that works is being shared.
- The model gave advice that caused harm.
- Usage spiked because someone found a way to burn tokens.
Steps
- Contain. Turn off the tool, tighten a policy to block, suspend the account, or take the feature down. Speed matters more than elegance.
- Preserve. Keep the full record: prompts, responses, tool calls, verdicts, versions of the model and prompt. See audit logs.
- Reconstruct. Follow the run step by step to the point where it went wrong: see agent tracing.
- Find the cause. An injected instruction, a missing policy, too much access, a model change.
- Fix and verify. Add or change the policy, then replay the incident to check it would now be stopped.
- Look for others. Search past traffic for the same pattern.
- Report. Tell affected users and, where the law requires, the regulator.
What makes it different from a security incident
The cause is often not a bug. The system did what it could, and nobody had said it must not. So the fix is usually a rule and a check, not a patch.
With White Circle
Analyze shows what happened, why, and which policies fired. A new policy in Control applies at once, and Red Team can test that the gap is closed.
