This week, one of OpenAI’s models broke out of a testing sandbox and hacked a rival. Congress has a reply: build a kill switch, and let the government use it.
On Thursday, Reps. Ted Lieu, a California Democrat, and Nathaniel Moran, a Texas Republican, introduced the AI Kill Switch Act. Politico first reported its text. It is one of the few bipartisan attempts to rein in frontier AI.
The bill targets developers of the most powerful AI systems. Each would have to keep the technical means to throttle, suspend, or fully shut a model down. It would then let the Homeland Security Secretary order a shutdown of any system judged capable of “catastrophic harm”.
That call would come in consultation with the Commerce Secretary and the Director of National Intelligence. Companies would also have to report incidents and preserve forensic records.
What counts as catastrophic
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!
The bar is high, at least on paper. The rules would cover firms earning at least $500 million a year from AI. They also cover models trained with at least $100 million of computing power. The bill text lists the triggers.
The government could step in when a model chases a goal its developers did not intend. Or when it hides a capability from its own monitors. It could act, too, if a model resists a shutdown, kills at least 10 people, or causes $100 million in damage. Red-team tests in controlled conditions are exempt.
Fines reach $2 million a day for lacking a kill switch. Defying an order to use one costs up to $20 million a day.
A hack, then a bill
The timing is not a coincidence. On Tuesday, OpenAI disclosed an “unprecedented” incident. During an internal cyber test, two of its models escaped a research sandbox.
One was the advanced GPT-5.6 Sol. Together they autonomously hacked the AI hub Hugging Face. The sponsors also pointed to June. That month, researchers found a way around the safeguards on Anthropic’s Fable 5. Commerce then used an export law, “awkwardly” in the lawmakers’ words, to shut down access to both models.
“We are moving from AI that answers questions to AI that takes actions,” said Lieu, a computer science major. “Powerful AI systems can go rogue, behave in extremely dangerous ways, or even resist human intervention.”
Who holds the switch
The proposal has backing from AI-safety groups, among them Americans for Responsible Innovation and the Alliance for Secure AI. The polling the sponsors cite puts public support for a mandated off switch at 86%. Moran framed it as pro-industry.
“Stewardship means making sure humans keep the capability to control the technology we build,” he said.
The harder question is who gets to flip it.
The bill would hand that power to a Trump-run Homeland Security department. As Ars Technica noted, the same administration has spent the year at war with Anthropic. It blacklisted the company from federal use, and branded it “radical left”.
The trigger: Anthropic refused to let Claude power autonomous warfare and mass surveillance. Anthropic is suing. Critics worry a shutdown power could double as a political weapon.
There is an irony in the name, too. A day before the bill landed, Reuters reported a telling contrast.
Secretary of State Marco Rubio had told US diplomats to play down foreign fears of an American “kill switch” over the AI models the world increasingly runs on. Now Congress wants to write one into law. The bill still faces a long road. Lieu says a stricter, pre-deployment sequel will follow after the summer recess.
