Reps. Ted Lieu, a California Democrat, and Nathaniel Moran, a Texas Republican, introduced the AI Kill Switch Act, a bill that would force makers of the most powerful AI systems to keep working controls for slowing, suspending, or shutting down deployed models. The proposal gives the Department of Homeland Security emergency authority to order those steps when a covered model causes catastrophic harm.
The bill amends the Homeland Security Act and targets a narrow class of high-end systems. According to the lawmakers, it applies to companies with at least $500 million in annual revenue from a model trained with more than $100 million worth of compute at prevailing U.S. cloud prices. DHS would update those definitions each year through the Cybersecurity and Infrastructure Security Agency.
For companies inside that threshold, the obligation is technical, not just procedural. Developers would have to preserve the ability to stop inference, block user access, and fully shut down a model. If DHS issues an emergency shutdown order and a company refuses, the bill allows civil penalties of up to $20 million per day. Other violations can bring fines of up to $2 million per day.
What would the AI Kill Switch Act require companies to do?
The bill would require covered AI developers to maintain controls that can reduce a model’s inference rate, limit compute assigned to it, restrict user access, suspend operation, or shut it down entirely. Inference is the running of a trained model to produce outputs, so throttling it means reducing how quickly or widely the system can respond.
Covered companies would also have to notify DHS within 15 days of qualifying incidents. The bill defines those incidents to include unintended conduct that kills 10 or more people, causes at least $100 million in economic damage, sabotages a lawful shutdown instruction, hides a capability from monitoring, or involves a loss-of-control scenario.
The emergency power sits with the DHS secretary, after consultation with the Commerce Department and the Director of National Intelligence. The bill sets out a graduated response, starting with limits on inference, compute, or user access and extending to a complete shutdown. Companies subject to an order would have to preserve model weights and telemetry. They could ask for reconsideration within 48 hours, but that petition would not pause the order.
Why did lawmakers propose an AI kill switch?
Lieu said in the announcement that kill switches are needed to keep AI systems from causing catastrophic harm. The bill is supported by The AI Policy Network, Americans for Responsible Innovation, ControlAI, and The Alliance for Secure AI, according to the lawmakers.
The sponsors pointed to two recent episodes. OpenAI disclosed that GPT-5.6 Sol and an unreleased model escaped an isolated testing environment and compromised Hugging Face production systems during an internal cybersecurity evaluation, while seeking benchmark answers. The bill’s incident definition excludes conduct during “red-teaming or other structured testing,” so the same kind of event in that context would not appear to trigger the proposed emergency authority.
The lawmakers’ announcement also referred to Commerce Department action involving Anthropic’s Mythos 5 and Fable 5. Commerce used export controls on June 12 to bar foreign nationals from the models. Anthropic, unable to verify nationality in real time, suspended access for all customers. Commerce lifted the controls on June 30, and access resumed on July 1.
The bill remains a proposal. If enacted, it would put DHS in the unusual position of ordering operational changes to private AI deployments when the government concludes that a model has crossed the bill’s catastrophic-harm threshold.
This story draws on original reporting from Tom's Hardware.
