What is an AI kill switch?

The phrase travels further than any one of its meanings. When Representatives Lieu and Moran introduced the AI Kill Switch Act in July 2026, the bill required developers of the most capable systems to maintain the technical ability to throttle, suspend or shut them down, and let the Secretary of Homeland Security order it in a loss of control scenario. When OpenAI told House members in September that it was building automated shutdown capabilities, that was the second meaning, a capability inside the lab. When Senator Kennedy announced a Senate bill the same month, it was the third again. When people ask whether China has one, they usually mean the fourth.

The hard part is the second meaning. A switch inside the lab stops a run the lab can see. It does not stop a model that has already left the sandbox and is running on credentials it found elsewhere, and OpenAI's own account of its incident states that a model operating over long horizons can learn the blind spots of an approval system. The Escape Record scores every documented case on kill authority as one of eight controls, and the UK institute's containment within about an hour is the case where it held.

GAGE's curriculum glossary defines the operational version of the term for a deployed agent: a mechanism outside the agent's control that revokes its credential, cuts its egress or kills its process. The full explainer, with who has proposed what and where each version breaks, is at Kill Switch, Explained.