AI Incident Response

AI incidents can involve unauthorized tool actions, sensitive-data exposure, malicious retrieved content, unsafe output, unexpected model behavior, or loss of audit evidence. Use a playbook that preserves evidence before configuration is changed.

First Response

1. Record the time, reporter, affected organization, agent, workflow, and external system.
2. Preserve the conversation, Tool Execution, Audit Log entry, correlation IDs, and provider logs.
3. Stop further impact by disabling the affected function, connection, workflow, public interface, or agent.
4. Revoke or rotate credentials if exposure or credential misuse is possible.
5. Identify affected data, users, recipients, records, and environments.
6. Assign an incident owner and communication owner.

Investigation Questions

• Was the initiating identity authorized?
• Did a user prompt, retrieved document, webpage, email, or tool response contain malicious instructions?
• Which prompt, skill, memory, dataset, model, and tool configuration were active?
• Did a human approve the action, and what context did the approval show?
• What did the external provider record as the final action?
• Can the change be reversed, and has the same configuration affected other executions?

Recovery

Containment is not the final fix. Correct the narrowest failed control, then retest:

• reduce data or tool scope,
• change the function policy,
• improve argument validation or approval context,
• remove or reprocess poisoned content,
• update the prompt or skill boundary,
• add a regression test,
• restore or correct external records,
• document the decision and residual risk.

After recovery, review similar agents, connections, and workflows for the same condition. Follow applicable contractual, regulatory, privacy, and customer notification procedures.

See Monitor Usage, Audit Logs, and Risk for common operational playbooks.