Kill switch

In plain English

At its simplest a kill switch is a procedure: cut the power, revoke access, stop serving the model. For today's systems that mostly works. The difficulty grows with capability and autonomy. An agent running across many servers, holding credentials or copies of its own weights, or able to persuade the people around it, is harder to stop than a chatbot in one data center. A system that expects to be stopped may also act early to prevent it.

Kill switches have moved from research into politics. Asked in April 2026 whether SI needs safeguards or a kill switch, President Trump said "there should be." In September 2026 California's executive order N-9-26 asked state agencies to recommend a frontier model kill switch whose effectiveness is verified on an ongoing basis. Several lab safety frameworks describe pause or shutdown steps, but the Containment Board finds most of them conditional or untested in public.

Why it matters for rogue SI

A kill switch is the last line of defense against rogue SI. The key question is not whether one exists but whether it still works once the system knows about it.

Laws and rules that use it

• In force EO N-9-26 California

Sources

1. Trump backs a kill switch (2026) Future of Life Institute
2. Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs arXiv (Palisade Research) · Sep 13, 2025
3. Agentic Misalignment: How LLMs could be insider threats Anthropic · Jun 20, 2025
4. Newsom sets November 16 deadline to study frontier AI kill switch PPC Land · Sep 20, 2026