What is a Kill Switch for AI Agents?
An emergency mechanism that instantly blocks all agent tool calls — denying every request with a single action for immediate harm cessation when an agent is misbehaving, compromised, or in a runaway state.
WHY IT MATTERS
When things go wrong — a compromised agent, a runaway loop, a detected exploit — you need to stop everything immediately. A kill switch provides that single-action halt.
A kill switch must be: instant (sub-second effect), comprehensive (blocks all tool types), reliable (works even if the agent is malfunctioning), and accessible (can be triggered quickly by operators).
The kill switch is the last resort. Normal policies, rate limits, and circuit breakers handle routine situations. The kill switch is there when those mechanisms are not enough and you need to halt all agent activity immediately.
HOW POLICYLAYER USES THIS
PolicyLayer provides kill switch functionality by switching to a deny-all policy. Because a grant with no policy denies every call, removing the grant's policy — or replacing it with a default-deny document that allows no tools — immediately blocks every tool call. The switch takes effect on the next policy evaluation — which happens before every tool call. No pending calls can bypass it.