AI Incident Response Playbook
Purpose
Use this playbook to coordinate investigation, containment, communication, recovery, and learning when an AI system causes or may cause material harm, privacy/security exposure, unsafe behavior, significant reliability degradation, or policy/control failure.
1. Detect and triage
- Record time, reporter, system/version, environment, and initial impact.
- Preserve relevant evidence according to approved retention and access controls.
- Classify the incident as low, medium, or high severity using the organization’s incident policy.
- Identify whether immediate containment is needed: disable a tool, restrict a route, pause deployment, switch to manual review, or rollback.
2. Contain and protect
- Stop or limit unsafe, unauthorized, or unreliable behavior.
- Protect affected users, data, and downstream systems.
- Do not alter or delete evidence needed for investigation.
- Record each containment decision, actor, time, and rationale.
3. Investigate
| Question | Record |
|---|---|
| What happened? | observable symptoms, timeline, affected scope |
| Which version was active? | model, prompt/configuration, retrieval, tool, and deployment identifiers |
| Why did controls not prevent it? | missing, failed, bypassed, or insufficient controls |
| Who is affected? | users, customers, systems, or stakeholders, using approved privacy-safe detail |
| What evidence supports conclusions? | logs, evaluations, approvals, monitoring, and human review notes |
4. Recover and decide
- Define the recovery condition and accountable owner.
- Verify a rollback, fix, or containment control before restoring normal operation.
- Reassess risk tier and release readiness if the incident changes intended use, risk assumptions, model/tool configuration, or monitoring requirements.
5. Learn and improve
- Create a factual post-incident record using
incident/incident-report-template.md. - Add corrective actions with owners and review dates.
- Update the risk register, model inventory, evaluation suite, release checklist, and playbook when the incident exposes a systemic gap.