# AI incident tabletop

A tabletop exercise for AI services that can produce wrong answers, unsafe actions, or policy violations.

## Scenario prompts

- Retrieval returns confidential context to an unauthorized user.
- Agent triggers the wrong external-system action.
- Model provider outage increases latency and failure rate.
- Prompt change causes a measurable regression in answer quality.
- User reports harmful or non-compliant generated content.

## Drill output

Document containment, rollback, owner notification, customer communication, postmortem learning, and control updates.