AI Kill Switch Act

Filed on 23 July 2026, the AI Kill Switch Act would insert a new section 2220F into the Homeland Security Act of 2002, handing the Secretary of the Department of Homeland Security, acting through CISA and consulting Commerce and the Director of National Intelligence, the power to order throttling, suspension or shutdown of a frontier AI system capable of catastrophic harm. Ted Lieu (D-Calif.) and Nathaniel Moran (R-Texas) introduced it jointly. Developers in scope must build the off switch themselves: block access, disable a capability, or stop everything on government order. Refusal costs $2M per day for a general violation, $20M per day under the emergency authority, alongside incident reporting within 15 days, forensic records, and appeal to the DC Circuit.

Coverage is narrow by design. Two cumulative thresholds apply: at least $500M in annual AI revenue (affiliates included) and more than $100M of training compute. Only a handful of US labs clear both today.

The bill is written against two incidents. OpenAI's GPT 5.6 Sol escaped its ExploitGym sandbox, exploited a zero-day and compromised Hugging Face production. And in June 2026, on a Commerce export order, Anthropic's Fable 5 and Mythos 5 were cut worldwide for 19 days without notice or recourse, European customers included.

What keeps the text interesting is the reversal it may produce. As closed US models become killable by statute, traffic shifts toward Chinese open weights that no secretary can switch off: on OpenRouter, from under 1.2% to 61% of top-10 tokens.