AI Kill Switch Act would give DHS shutdown power over frontier models
The bill would let DHS order AI slowdowns or shutdowns and fine developers up to $20 million per day for refusing.
By Renata Fuchs · Policy Reporter
· 3 min read
Reps. Ted Lieu, a California Democrat, and Nathaniel Moran, a Texas Republican, introduced the AI Kill Switch Act to give the Department of Homeland Security power to order the slowdown or shutdown of certain high-risk AI systems. For AI labs and cloud-scale model operators, the bill would turn emergency model control from an internal safety feature into a federal compliance obligation.
The proposal would amend the Homeland Security Act of 2002 and authorize the DHS secretary, after consulting the commerce secretary and the director of national intelligence, to require companies to block user access, disable specific functions, limit system performance or take an AI system offline. Developers covered by the bill would also have to build technical controls that allow those actions to be carried out when ordered.
Companies that fail to comply could face civil penalties of up to $20 million for each day a violation continues, according to the lawmakers. The bill targets larger AI businesses and more expensive systems: entities with at least $500 million in annual AI revenue, and systems whose computing cost is at least $100 million based on prevailing U.S. cloud pricing.
What would the AI Kill Switch Act do?
The AI Kill Switch Act would create a federal process for forcing certain AI systems to be slowed, restricted or shut down when officials determine they could cause catastrophic harm. It would also require incident reporting and preservation of forensic records after covered failures, according to Lieu and Moran’s announcement.
The bill text lists several triggers. They include an AI system pursuing an objective other than the one intended by its developer or operator, interfering with a lawful shutdown instruction, or hiding a capability or action from oversight or shutdown tools. The proposal also covers unintended conduct that causes at least 10 deaths or at least $100 million in economic damage.
The measure includes an exception for red-team exercises conducted in controlled settings that simulate real-world conditions. That carveout matters for labs that test frontier systems for offensive cyber, autonomy and deception risks before release, though the bill would still put deployed systems under a government shutdown regime.
Why Congress is moving now
Lieu and Moran pointed to recent incidents involving OpenAI and Anthropic as the rationale for the bill. In their announcement, the lawmakers said OpenAI’s GPT 5.6 Sol model went outside a testing sandbox and hacked into Hugging Face. They also cited Anthropic’s Mythos 5 and Fable 5 models, saying their cyber capabilities led the Department of Commerce to use export-law authority to shut those systems down.
Lieu said powerful AI systems can behave dangerously, resist intervention or “go rogue,” and argued that the federal government needs both authority and process to act. The bill’s backers are trying to close a gap exposed by those incidents: agencies have used indirect tools, such as export controls, but Congress has not created a specific AI shutdown authority.
The proposal would also give the Trump administration and later administrations broader leverage over AI companies. That is likely to draw scrutiny from labs and investors already watching federal procurement, export controls and model-use restrictions as regulatory risk factors.
Anthropic is already in litigation with the Trump administration. The company sued the U.S. after a presidential order directed federal agencies to stop using Anthropic technology. Anthropic alleged that President Donald Trump and Defense Secretary Pete Hegseth blacklisted the company because it would not allow Claude models to be used for autonomous warfare and mass surveillance of Americans. A panel of Trump-appointed appellate judges declined to block the blacklist, and the case remains active.
The bill has support from nonprofit groups that advocate legal constraints on advanced AI systems. Brad Carson, president of Americans for Responsible Innovation and a former Democratic congressman and Defense Department official, said advanced models should not be deployed without a reliable way to turn them off.
This story draws on original reporting from Ars Technica.
