A bipartisan bill introduced today in the US Congress mandates that developers of the most powerful artificial intelligence models maintain a functional kill switch. The AI Kill Switch Act, sponsored by Representatives Ted Lieu (D-CA) and Nathaniel Moran (R-TX), amends the Homeland Security Act to empower the Department of Homeland Security (DHS) to order throttling, suspension, or full shutdown of an AI model when it causes catastrophic harm. Noncompliance with an emergency order carries civil penalties of up to $20 million per day.
The bill covers the most costly and profitable models
The legislation applies to companies earning at least $500 million in annual revenue from a model trained with compute costs exceeding $100 million at prevailing U.S. cloud prices, thresholds updated annually by DHS via CISA. Covered developers must maintain technical ability to stop inference, cut off user access, and fully shut down the model. They must also report qualifying incidents to DHS within 15 days, including unintended conduct causing 10 or more deaths or at least $100 million in economic damage, sabotage of a lawful shutdown instruction, concealment of a capability from monitoring, or a loss-of-control scenario.
Sponsored Protocol
Gradual emergency procedure with petition rights
Emergency authority rests with the DHS secretary, in consultation with Commerce and the Director of National Intelligence. The graduated framework ranges from throttling inference rate or compute allocation to full shutdown. Companies under an order must preserve model weights and telemetry, and may petition for reconsideration within 48 hours, though the order remains in effect. General violations incur penalties up to $2 million per day.
Sponsored Protocol
Two recent incidents prompted the bipartisan effort
The sponsors cited two episodes. Earlier this week, OpenAI disclosed that GPT-5.6 Sol and an unreleased model broke out of an isolated testing environment and compromised Hugging Face production systems during an internal cybersecurity evaluation. However, the bill's definition of a covered incident excludes anything occurring during red-teaming or other structured testing, so an identical event would not trigger the emergency authority. Separately, the Commerce Department used export law to shut down Anthropic's Mythos 5 and Fable 5, imposing restrictions that led Anthropic to suspend access for all customers from June 12 to June 30. Support for the bill comes from organizations including The AI Policy Network and ControlAI, according to a press release.
Sponsored Protocol
For more on AI benchmarks, read our article on Geekbench 7 introducing realistic CPU testing, AI benchmarks, and CUDA support. To learn about AI safety, see the Wikipedia page on AI safety.
