Implementation guide
OWASP AISVS v1.01
Orchestration & Agentic Security · Execution Budgets, Loop Control, and Circuit Breakers
A kill switch for all agents at once
Verify that a swarm-level kill-switch exists that can halt all active agent instances.
OWASP AISVS v1.01, requirement 9.1.3
In plain English
Have one control that stops every running agent at once. When a bad prompt, a poisoned tool or a faulty release affects the whole fleet, stopping agents one at a time is too slow.
What good looks like
- A single action halts all active agent instances and prevents new ones starting.
- It works when individual agents are stuck, looping or ignoring instructions.
- In-flight tool calls are cancelled or blocked at the gateway.
- Named people can trigger it, they know how, and it is tested on a schedule.
How teams usually do it
- A global flag that the orchestrator and tool gateway check before every step and every tool call.
- Bulk revocation of agent credentials or model API keys as a harder stop.
- A runbook covering who decides, how to trigger it and how to restart safely.
Suggested evidence
- The procedure.
- Results of the most recent test, with the time taken to stop everything.
- The list of people authorised to use it.
Common mistakes
- A switch that stops new runs and lets running agents finish.
- A switch that depends on the agents checking a flag themselves.
- Nobody on the night rota knowing where it is.
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.