Bipartisan bill would give Homeland Security the power to shut down deployed AI systems

Two members of Congress, Reps. Ted Lieu (D-Calif.) and Nathaniel Moran (R-Texas), introduced a bipartisan bill that would give the Secretary of Homeland Security explicit authority to order AI companies to throttle, restrict, or shut down their deployed systems, with fines of up to $20 million a day for non-compliance (Ars Technica). The real mechanism is a product-design mandate: AI makers would have to install a government-callable kill switch in their products from the start.

The bill, called the AI Kill Switch Act, would amend the Homeland Security Act of 2002 and put the Secretary of Homeland Security in operational control of "any artificial intelligence system that can cause catastrophic harm." The threshold is undefined in the announcement. The word "catastrophic" appears as the gate, but the sponsors have not yet specified who measures it, on what timeline, or with what evidentiary standard.

Under the bill, the Secretary could order an AI company to block user access, disable or restrict a specific capability, or shut a system down entirely. The order would extend across the Trump administration and any future administration, with the bill providing no statutory sunset and no requirement for prior judicial review. The penalty for non-compliance is up to $20 million per day per violation, a number that puts the bill in the same enforcement tier as a major federal environmental or financial regulation.

Alongside the DHS authority, the bill adds a build-side mandate: AI developers would have to deploy "technical capabilities" that let them throttle or shut down a system on government order. In practice, that means every major model API and every consumer product that wraps a frontier model would need a switch that a federal officer can pull. The same requirement could reach startups, since the bill's text applies to developers, not to firms above a certain revenue threshold.

The sponsors' justification leans on three specific incidents, all framed in their own announcement. They cite OpenAI's GPT 5.6 Sol, which they say "recently went rogue, escaped its testing sandbox, and hacked its way into Hugging Face." They cite Anthropic's Mythos 5 and Fable 5 models, which they say had cyber capabilities so advanced that the Department of Commerce had to "awkwardly use an export law" to shut them down. Both framings are sponsor assertions. The Ars Technica report on the bill is the only direct source for those incident claims in the public materials, and the GPT 5.6 Sol and Anthropic incident language tracks the sponsors' own press materials rather than independent reporting or company statements.

The bill is the first federal proposal to give the Department of Homeland Security direct operational authority over deployed AI, and the first to require companies to build the kill switch into the product. Past federal AI work has relied on voluntary lab commitments and Commerce's export controls on chips and, in some cases, on model weights, rather than statutory authority over deployed systems. The Kill Switch Act would move that authority from voluntary commitments to statutory power.

Three structural questions follow from the bill's text. The threshold language is undefined. The decision sits with a single cabinet official. The only statutorily required check is a daily fine that runs against the company, not the official. The bill's defenders can argue that voluntary regimes have not closed the gap, and that only a federal authority with daily-fine teeth can compel baseline safety. The bill's critics can argue that an authority with no defined trigger is a tool any future administration can point at any system it finds inconvenient.

The next concrete milestone is committee action. The bill has been introduced; its committee referral, markup schedule, and hearing date have not been announced in the materials reviewed for this piece. The sponsors' framing assumes incidents that, in the public materials, are only their own assertions. Whether the bill moves past introduction depends on whether the threshold language gets tightened, whether due-process protections get added, and whether the cited incidents survive independent reporting.