The off switch: who can stop an AI system, and how

AI agents now act in seconds, and the controls meant to stop them are patchy, voluntary and rarely tested. Where an off switch can sit, which ones work, who holds them today, and what a working one needs.

Share

Updated 28 Oct 2026 · By Angela Milash

Start here, in plain words

AI systems no longer just answer questions. They take actions: they send money, change records, write code, and in war, help choose targets. When one of them goes wrong, someone has to be able to stop it, fast.

That someone is usually assumed to exist. Often it does not. The company that built the AI model points to the customer. The customer points to the vendor. The engineers point to their managers. And the people with real power to say no, governments and boards, are often the slowest to act.

This paper explains what an off switch is, where it can sit, which ones work, who holds them today, and what a working one needs. It ends with questions you can take into your next board or security meeting.