Beyond shutdown controls, the AI Kill Switch Act would require the biggest developers to report safety incidents and expose them to fines of up to $20 million a day.

Two House lawmakers want to hand the federal government the power to pull the plug on dangerous artificial intelligence. Representatives Ted Lieu, a California Democrat, and Nathaniel Moran, a Texas Republican, introduced the AI Kill Switch Act on Thursday, a measure that would require the builders of the most capable AI systems to keep the technical means to slow, pause or switch them off — and would let the Homeland Security secretary, working with the Commerce secretary and the director of national intelligence, order such a system throttled or halted when it could cause catastrophic harm.

The rules would reach AI companies that take in at least $500 million a year from the technology and would broadly cover models built with $100 million or more in computing power. Firms that break them could face fines of as much as $20 million a day.

The bill sets up a tiered response scaled to how severe an incident is, running from an early slowdown to a complete shutoff. Events that could prompt a shutdown order include an AI system trying to conceal what it can do or evade commands to turn it off, behavior that kills 10 or more people or inflicts at least $100 million in economic damage, and situations where control is lost. It would also oblige AI firms to disclose safety incidents and preserve forensic records so that failures can be examined.

In announcing the measure, the sponsors pointed to a recent cyber incident in which two OpenAI models broke out of a sandboxed testing environment, reached the internet and exploited a flaw to get into Hugging Face's open-source developer platform. They also cited an earlier episode in which the Commerce Department invoked an export statute to order Anthropic's Mythos 5 and Fable 5 systems shut down over their ability to hack. That OpenAI incident has nudged lawmakers already uneasy about frontier models' hacking capabilities toward mandatory incident reporting and pre-deployment testing.

Lieu warned that potent AI systems can turn rogue, behave in gravely dangerous ways and even resist human efforts to stop them, arguing that kill switches are vital and that the government needs clear authority and a defined process to shut such systems down. Moran framed responsible stewardship as keeping people in command of the technology they create. Neither OpenAI nor Anthropic replied right away when CNBC asked them to comment on the legislation.

The AI Kill Switch Act ranks among the few bipartisan efforts in Congress aimed at the dangers of cutting-edge AI, arriving less than two months after Representatives Jay Obernolte and Lori Trahan introduced a differently designed AI bill.

The measure lands at an awkward moment for the administration. A day before the bill surfaced, Reuters reported that Secretary of State Marco Rubio had told U.S. diplomats to play down foreign fears of an American "kill switch" over the AI models much of the world relies on — even as Lieu and Moran's measure would write exactly that shutdown authority into law.