AI Incident Readiness
Three AI Incident Tabletop Scenarios That Test Executive Decisions
Plan an executive AI tabletop around decision authority, evidence and recovery, using three clearly hypothetical scenarios.
Top One Percent AI · Practical guidance with linked sources.
An AI incident tabletop should expose a decision your organization has not settled. Who can pause an agent? Who verifies an unusual payment request? Who decides whether an AI-generated answer has created a customer issue?
A dramatic scenario can start a conversation. A useful exercise leaves the team with named decisions, owners and work to complete.
The three scenarios below are hypothetical planning examples, not accounts of actual incidents or a complete response plan. Choose one that resembles a workflow your organization uses. Have your security, legal, privacy and business owners adapt it before the exercise.
Scenario one: an agent acts beyond its brief
A customer-service agent is allowed to retrieve order information and draft responses. During the exercise, participants learn that it has also changed records after processing an external document. The team does not yet know which changes were authorized.
This scenario tests the difference between generating an answer and taking an action. OWASP's excessive-agency guidance identifies unnecessary functionality, permissions and autonomy as sources of risk.
Ask the room:
- Who can suspend the agent's ability to make changes without shutting down customer support entirely?
- Which business owner decides what service can continue manually?
- What records would show the tool calls, affected accounts and approvals?
- Who decides when the agent can resume work, and what evidence must they see?
Do not accept “IT will handle it” as the final answer. Record the role, the authority and the unresolved dependency.
Scenario two: a convincing executive request
A finance employee receives a voice message that sounds like a senior executive. A follow-up message asks for an urgent change to payment instructions. The exercise begins before any payment is made.
The joint NSA, FBI and CISA deepfake advisory describes synthetic-media threats that impersonate leaders and use fraudulent communications. The planning question is how your organization verifies authority when voice or appearance is insufficient.
Ask who owns a trusted callback route, who can hold the transaction and how an employee escalates the request without being penalized for delaying an executive. Then introduce a second message claiming that the normal approver is unavailable. Does the verification process survive the pressure?
Scenario three: sensitive information in an AI response
An employee reports that an internal assistant returned information from a document they should not have been able to access. The team has a screenshot, but it does not yet know whether the problem affects other users.
Ask who gathers the relevant access and application records, who determines the potentially affected population, and who coordinates with the service provider. Separate confirmed exposure from assumptions. Decide who can restrict the affected feature while preserving information needed for investigation.
Avoid putting real confidential documents into the exercise. Use synthetic records clearly marked for training.
End with decisions, not a participation score
NIST SP 800-61 Revision 3 places incident response within broader cybersecurity risk management. That is a useful frame for the debrief: the exercise should feed preparation and improvement, not sit apart from them.
For each gap, capture the decision, accountable owner, evidence needed and review date. Distinguish a discussion agreement from a control that has been implemented and tested.
When selecting a speaker or facilitator, ask which of these decisions the engagement will exercise. Confirm preparation, participant roles, written findings and follow-up separately. A keynote can introduce the issues; a tabletop needs participation and a defined scope.
Compare session formats or build a high-level event brief before requesting an engagement.