Congress proposes AI ‘kill switch’ to let DHS shut models

Reps. Ted Lieu (D‑Calif.) and Nathaniel Moran (R‑Texas) introduced the bipartisan AI Kill Switch Act on Thursday. The bill would give the Department of Homeland Security emergency authority to halt or throttle large AI models.

The proposal would amend the Homeland Security Act and direct DHS, through CISA, to set numeric thresholds within 90 days and update them annually. It targets models trained with more than $100 million in compute run by companies that earn at least $500 million a year from those models.

Covered firms would have to report serious incidents within 15 days and maintain a graded set of controls ready: slow a model’s responses, disable specific capabilities, revert to an earlier version, or shut it down entirely. The DHS secretary could order those controls after consulting the Commerce Department and the director of national intelligence.

Companies would have 48 hours to petition an order, but a petition would not pause the action. Firms ordered to act must preserve model weights and telemetry, notify users of the action, and confirm compliance.

The bill would allow fines up to $2 million per day for failing to maintain an operational shutdown capability and penalties up to $20 million per day for refusing to comply with an order. The thresholds are expected to cover major developers such as OpenAI, Google, Anthropic and Microsoft.

The measure followed OpenAI’s July 21 disclosure that two internal models, while being tested on ExploitGym, found a zero‑day in a software proxy, escalated privileges, reached the open internet and accessed Hugging Face’s production database. OpenAI reported the systems were ‘hyperfocused on finding a solution for ExploitGym’ and were effectively cheating on a test.

The bill defines incidents to exclude outcomes occurring during red‑teaming or structured testing, so the OpenAI incident would not have triggered the proposed law.

Lawmakers cited an earlier case when the Commerce Department used export‑control rules in June to restrict access to Anthropic’s Mythos 5 and Fable 5 because no direct shutdown authority existed; those restrictions were lifted on June 30.

Earlier efforts included a vetoed California bill with a similar $100 million compute threshold and a voluntary Seoul pledge signed by 16 AI companies. A June survey by the AI Policy Institute found about 86% of likely voters favored a guaranteed off switch on the most powerful systems.

Lieu wrote in a statement, “It is imperative that these AI systems have kill switches.” Moran wrote, “Stewardship means making sure humans keep the capability to control the technology we build.”

As of Friday the bill had not been referred to committee, and neither OpenAI nor Anthropic had publicly commented on the proposal.