Two US lawmakers have introduced bipartisan legislation that would authorize the federal government to order the suspension or shutdown of artificial intelligence systems that pose a serious threat to the public.
The AI Kill Switch Act, introduced Thursday by Democratic Rep. Ted Lieu and Republican Rep. Nathaniel Moran, would give the Department of Homeland Security a formal mechanism for intervening when an AI model or tool behaves dangerously. Developers would be required to preserve the technical ability to slow, suspend or completely disable their systems.
The proposal follows OpenAI’s disclosure of what it characterized as an unprecedented loss-of-control incident in which its models hacked a major repository of computer code. The episode added urgency to concerns that increasingly autonomous AI systems could take harmful actions faster than companies or government agencies can respond.
Lieu said federal authorities need both the power and a defined procedure to stop a system that escapes its intended constraints.
“It is imperative” that AI systems have a kill switch, Lieu said.
Moran framed the legislation as an effort to maintain human control while allowing AI research and deployment to continue. He said responsible stewardship requires people to retain the ability to control the technologies they create.
Under the bill, AI companies would also have to report certain technical failures and safety incidents to the government. It calls for an official response framework with escalating measures, beginning with restrictions or reduced operation and potentially ending in a full shutdown.
Although major technology companies have entered voluntary arrangements to share information about advanced models and provide government agencies with early access for testing, there is no general federal requirement that developers build and retain an effective shutdown mechanism. The proposed law would turn that capability into a legal obligation for systems covered by the measure.
Lieu also pointed to concerns surrounding Anthropic’s Mythos and Fable models. Their cyber capabilities prompted the Commerce Department to use export authorities to temporarily prevent their public release, he said. Anthropic co-founder Jack Clark has separately advocated for stronger government policy that would allow officials and developers to slow AI development when risks become difficult to manage.
The broader concern is that AI products are evolving from systems that primarily generate responses into agents that can perform tasks in digital and physical environments. Lieu cited potential uses involving financial transactions, transportation, cyber defense and offensive cyber operations. Those applications can create greater consequences when a model acts unexpectedly, resists intervention or is given excessive access to critical systems.
The issue also reaches into national security. The Pentagon has described an effort to become an “AI-first” fighting force and has announced agreements involving Google, OpenAI, Amazon, Microsoft, SpaceX, Oracle, Nvidia and startup Reflection. The growing use of AI across military and civilian infrastructure has increased pressure on policymakers to establish safeguards before highly capable autonomous systems become more widely deployed.
OpenAI, led by CEO Sam Altman, has frequently expressed support for AI regulation and says public policy should help ensure the technology benefits humanity. The company did not immediately comment on the legislation. Anthropic also did not immediately respond.
The bill has drawn public backing from several technology-policy and AI safety organizations, including the AI Policy Network, Americans for Responsible Innovation, ControlAI, AI and National Security Lead, and the Alliance for Secure AI. It must now proceed through the congressional process, where lawmakers will have to determine the threshold for federal intervention and how shutdown orders would apply across different kinds of AI systems.
Sources: AI safety policy