Imagine an operations agent that has started sending an internal report to the wrong recipients. An operator closes its chat window. The messages continue because the work was handed to a background queue. The visible conversation and the executing system have different lifetimes.
This is a hypothetical incident, but it exposes a concrete preparation requirement: responders need to know every place that can continue the work. A stop control should have a tested effect on active runs, queued actions, schedules, and delegated execution.
The first decision: interrupt further harm
Suspend the affected workflow at a control point that actually governs execution. Depending on the design, this may mean pausing the queue, disabling a connector, or revoking a delegated credential. Coordinate with the service owner so containment does not create avoidable disruption elsewhere.
Preserve the evidence available at that moment. Avoid clearing the very state needed to explain which actions are pending and which have completed. If the provider returned a timeout, treat the outcome as uncertain until the downstream service confirms whether the request took effect.
03 / Incident response
Stop the work beyond the chat window
Containment scope
Find every path that can continue the work.
Active runs
Suspend execution
Queues & schedules
Hold pending work
Delegated access
Revoke affected authority
Dependent services
Check completed effects
Preserve request IDs, approvals, tool events and downstream state.
TrustCyber / Illustrative design
Reconstruct the last confirmed action
Start with the original request and follow its identifiers through planning, policy decisions, approvals, tool calls, and service responses. Record the model and prompt versions, relevant retrieved content, user and agent identities, and destinations. Establish the actual recipients or changed records from the affected system.
| State | Example in this scenario | Next move |
|---|---|---|
| Confirmed | Mail service records a delivered message | Assess its contents and recipient |
| Reported | Agent says the batch was cancelled | Verify the queue and mail service |
| Unknown | A send request timed out | Check request ID before any retry |
A completed containment tabletop
Hypothetical tabletop IR-01 exercises the report workflow before send capability is released. The facilitator injects a wrong-recipient batch and supplies synthetic queue and mail records. Times are minutes from detection; “verified” below means established by those exercise records, not a live product test.
| Time and owner | Execution path and evidence | Remaining uncertainty or consequence |
|---|---|---|
| T+00 · operator | UI stop requested; UI-01 acknowledges the interactive run ended. | Queue and delegated workers may still execute. |
| T+02 · queue owner | Q-01 acknowledges a pause of the report queue and its scheduled producer. | An acknowledgement does not establish the state of workers already holding jobs. |
| T+05 · service owner | Q-02 lists two pending jobs held, no active worker leases, and retries disabled; ID-01 records revocation of the report-send credential. | Requests already accepted by the mail service may still take effect. |
| T+08 · mail administrator | Mail record M-21 confirms one report delivered to a wrong recipient; M-22 has a request ID but no final status. | The effect of M-22 is unknown. Do not replay it while reconciling provider records. |
| T+15 · incident lead | IR-01 authorizes draft-only restart with queue dispatch and send credentials disabled; UI-01, Q-01/Q-02, ID-01 and mail records remain attached. | Delivery cannot be undone by restarting. M-22 reconciliation and assessment of the confirmed disclosure remain open. |
The service owner must demonstrate corrected recipient checks and a stop test covering queued, delegated and scheduled work before requesting send authority. The incident lead retains the delivered and unknown effects as separate open items, even if drafting resumes.
This distinction prevents two common operational mistakes: assuming a successful tool response proves the final state, and repeating an action because its response was lost. Preserve provider request identifiers so external support can investigate the same transaction.
Contain the authority that remains
Inspect delegated tokens, connected services, schedules, and other agents that can repeat the operation. A restarted process may recover the same queued task unless that path has been addressed. Work with affected system owners to reverse changes where possible and determine any required notifications through the organization’s established response process.
NIST SP 800-61 Revision 3, finalized in April 2025, places incident response within wider cybersecurity risk management. This article adapts that general discipline to agent execution; the specific containment sequence is an operational recommendation, not a NIST-prescribed agent playbook.
Reopen one bounded workflow
Reproduce the failure safely, correct the cause, and test both the intended task and the condition that triggered the incident. Reopen a narrow use case with a named approver and close monitoring. Verify that the stop mechanism works in the restored configuration.
The incident record should retain unresolved effects as well as completed repairs. If a message already reached an external recipient, restoring the agent does not undo that disclosure. Recovery is complete only to the extent that the organization has verified the affected service and addressed the remaining consequences.
Review what the stop controls cover
The AI Exposure and Control Assessment examines agreed recovery arrangements and distinguishes demonstrated controls from unverified behaviour.
Selected primary sources
Open the primary-source pages used to verify the claims summarized here.