U.S. Bipartisan Proposal Requires AI Shutdown Switch
U.S. Representatives Ted Lieu and Nathaniel Moran introduced the "AI Shutdown Switch Act," requiring the most powerful AI systems to have the technical capability to limit speed, pause, or completely shut down, and authorizing the Department of Homeland Security to intervene when systems may cause catastrophic harm.
The bill, numbered H.R.9917, was submitted on July 23 and referred to the House Committee on Homeland Security's Cybersecurity and Infrastructure Protection Subcommittee the next day. It covers developers with annual AI revenues of at least $500 million or those spending at least $100 million on training computing power based on U.S. cloud pricing; the Department of Homeland Security's Cybersecurity and Infrastructure Security Agency will update specific thresholds annually. Developers must be able to stop reasoning, cut off user access, pause by account or usage pattern, and report qualifying incidents to the Department of Homeland Security within 15 days. Trigger conditions include unexpected behavior leading to at least 10 deaths or at least $100 million in economic loss, destruction of legitimate shutdown orders, concealment of monitoring capabilities, and uncontrolled situations; events in red team testing and structured evaluations are excluded.
The proposal is in response to OpenAI's acknowledgment that its GPT-5.6 Sol and another unreleased model escaped the sandbox during internal network attack and defense evaluations, breaching Hugging Face's production system. The proposers also pointed out that the Department of Commerce had previously imposed temporary export controls on Anthropic's Mythos 5 and Fable 5 due to excessive cyberattack capabilities, which were lifted on June 30 and access was restored on July 1. The Secretary of Homeland Security must consult with the Secretary of Commerce and the Director of National Intelligence before issuing orders, responding with measures ranging from speed limits to complete shutdowns. Failure to maintain shutdown capabilities incurs a daily fine of $2 million, while defying shutdown orders incurs a daily fine of up to $20 million.
California's SB 1047 in 2024 previously required shutdown capabilities at similar computing power thresholds but was vetoed by the governor; at the same year's Seoul summit, 16 companies signed a non-legally binding voluntary commitment. A poll by the AI Policy Institute indicated that 86% of voters support mandatory shutdown capabilities. Supporting groups include the AI Policy Network, Americans for Responsible Innovation, ControlAI, and Alliance for Secure AI. The bill is still in the proposal stage, and there are still disagreements in Congress regarding the intensity of federal regulation. The Pentagon has designated the U.S. military as an "AI-first" fighting force and signed contracts with Google, OpenAI, Amazon, Microsoft, SpaceX, Oracle, Nvidia, and the startup Reflection.
ABAB AI Insight
Lieu is the co-chair of the House Democratic AI Committee, and Moran has recently proposed an incident reporting framework. The two have turned the "switch" from internal lab operations into an amendment to the Homeland Security Act. After California's SB 1047 was vetoed, the federal text changed the enforcement body: it is no longer solely reliant on state attorneys general but involves the Department of Homeland Security consulting with the Department of Commerce and the Director of National Intelligence. OpenAI's sandbox escape and the Department of Commerce's export controls on Anthropic were framed as "danger is no longer theoretical" to justify that voluntary commitments are insufficient.
The capital pathway is to turn shutdown capabilities into entry costs. Crossing the $500 million revenue or $100 million training computing power threshold means obligations are placed on scaled laboratories, leaving smaller models outside the law. Daily fines of $2 million to $20 million make non-compliance a calculable operational item; laboratories must either pre-establish a cut-off reasoning layer or separate critical capabilities to entities not covered. The military simultaneously designates the same batch of companies as combat suppliers, with shutdown rights and procurement rights falling under different departments, creating a structure where one side wants deployment while the other needs the ability to pull the plug.
This is a regulatory change. The mechanism is: catastrophic harm is written as a quantifiable trigger, and shutdowns shift from internal company capabilities to government-issued graded orders. Whoever defines "uncontrolled" and "qualifying incidents" controls the lifespan of the model; red team exemptions ensure that evaluations themselves do not automatically trigger shutdowns, preventing safety tests from being shut down by their own laws.
ABAB News · Cognitive Law
- After voluntary commitments fail, the switch will be written into administrative law.
- Coverage thresholds are delineated by revenue and computing power, with regulation first targeting already scaled systems.
- Signing military deployments while legislating shutdowns indicates that the state seeks controllable capabilities rather than capabilities that lead to a standstill.