# What is AI incident response?

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](/glossary/jailbreak) that works is being shared.
- The model gave advice that caused harm.
- Usage spiked because someone found a way to burn tokens.

## Steps

1. **Contain.** Turn off the tool, tighten a policy to block, suspend the account, or take the feature down. Speed matters more than elegance.
2. **Preserve.** Keep the full record: prompts, responses, tool calls, verdicts, versions of the model and prompt. See [audit logs](/glossary/audit-logs).
3. **Reconstruct.** Follow the run step by step to the point where it went wrong: see [agent tracing](/glossary/agent-tracing).
4. **Find the cause.** An injected instruction, a missing policy, too much access, a model change.
5. **Fix and verify.** Add or change the policy, then replay the incident to check it would now be stopped.
6. **Look for others.** Search past traffic for the same pattern.
7. **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](/products/analyze) shows what happened, why, and which policies fired. A new policy in [Control](/products/control) applies at once, and [Red Team](/products/red-team) can test that the gap is closed.

