Kill switch

A mechanism that stops an agent when other controls have failed, which must live outside the agent's control and must not depend on the agent's cooperation. Real kill switches act on the infrastructure beneath the agent: revoke its credential, cut its egress or disable its tools at the platform, kill its process or container, or freeze the account it acts through.

An automatic or manual control that stops an agent or an agent system. A layered design includes an automatic pause on a soft-limit breach, an automatic halt on a hard-limit breach requiring human authorization to restart, and an emergency shutdown for critical anomalies. Distinct from monitoring, which informs a human rather than acting.

The ability to interrupt or halt an AI system entirely, required for high-risk systems by EU AI Act Article 14; the top of the escalation path, held by a named, authorized person, and for the most severe systems paired with the four-eyes principle (two independent humans confirming) so the stop decision itself is robust.

A simple, one-move way to immediately stop an agent and route its work back to humans; the control that keeps an owner in charge of an autonomous worker.

The ability to disable one specific capability of a live system fast, without taking down the whole service, the moment a mitigation is shown to have failed in production. The last line of defense behind every other containment layer, and only useful if you know, before launch, which capability you would disable first and how long that takes.