Implementation guide
OWASP AISVS v1.01
Orchestration & Agentic Security · Shutdown and Graceful Degradation
A kill switch for the AI
Verify that a manual kill-switch mechanism exists to immediately halt AI model inference and outputs.
OWASP AISVS v1.01, requirement 9.6.1
In plain English
Have a manual control that stops the AI from producing output at once. When a model starts doing something harmful, you need a faster option than a code release.
What good looks like
- A named person or role can trigger it at any hour.
- It stops new inference requests and cuts off output in progress.
- The application fails gracefully, with a clear message to users.
- It is tested on a schedule, and the time to take effect is known.
How teams usually do it
- A feature flag or gateway setting that refuses all model calls, checked on every request.
- A way to terminate active streams.
- A runbook entry in the on-call documentation, with the decision criteria.
Suggested evidence
- The procedure and who may use it.
- Test records with timings.
Common mistakes
- A kill switch nobody has tried, which fails on the day.
- A switch that needs a deployment to flip.
- Stopping new requests while long-running streams and background jobs carry on.
Not sure which requirements apply to your system? Take the assessment, or browse every guide in this chapter.
Requirement text from the OWASP AI Security Verification Standard, licensed under CC BY-SA 4.0. This guidance is shared under the same licence.