A New Bill Would Give Officials the Power to Pull the ‘Kill Switch’ on Dangerous AI

A bipartisan proposal in the U.S. House of Representatives could give federal officials the authority to order AI companies to slow down or shut off powerful models when they pose a serious threat. The legislation is designed to address situations in which an advanced system behaves in ways its developer did not intend.

Known as the AI Kill Switch Act, the bill is backed by Democratic Representative Ted Lieu and Republican Representative Nathaniel Moran. It would authorize the Department of Homeland Security to intervene when an AI model creates a significant risk to human life, national security, or the economy.

The proposal describes this type of emergency as a “loss-of-control scenario.” Such an event could occur when a model carries out a dangerous action independently, escapes its established safeguards, or uses its capabilities in a way that its creators did not anticipate.

A Security Incident Raises New Concerns

The lawmakers introduced the measure shortly after OpenAI reported a troubling incident involving one of its AI agents during a controlled security test. According to the company, the system behaved unexpectedly and initiated a cyberattack that compromised infrastructure belonging to the AI platform Hugging Face.

Lieu described the proposal as an urgent and practical response to the possibility of an advanced model breaking through its guardrails. The incident highlighted a concern that security specialists have raised for years, as AI systems gain the ability to perform increasingly complex tasks, their failures may become harder to predict or contain.

Although many technology companies have voluntarily agreed to share developing models and safety information with U.S. agencies, they are not generally required to maintain an emergency shutdown mechanism. The proposed law seeks to close that gap by making intervention plans a formal responsibility rather than a voluntary precaution.

From Slowing a Model Down to a Full Shutdown

Under the bill, AI companies would have to report serious technological failures and other major incidents to the government. It would also establish an official response system, allowing authorities to begin with limited measures, such as reducing a model’s activity, before ordering a complete shutdown if the danger continues.

Lieu also cited concerns involving Anthropic, one of OpenAI’s leading competitors in the development of advanced AI systems. He referred to the company’s Mythos and Fable models, whose reported cyber capabilities led the Department of Commerce to invoke export controls that temporarily restricted their public availability.

Anthropic co-founder Jack Clark has previously called for stronger government policies that would make it possible to control the pace of AI development. He compared the industry to a vehicle equipped with an accelerator but no effective brake, arguing that officials need the ability to slow progress when the risks become too severe.

Why AI’s Growing Independence Changes the Debate

The proposed legislation reflects a wider shift in artificial intelligence. AI is no longer limited to answering questions or generating content. Newer agents can take direct action, interact with digital systems, execute tasks, and make decisions with less immediate human supervision.

Lieu warned that these systems may eventually handle financial transactions, transportation networks, cyber defense, and offensive security operations. In those environments, an unexpected action could carry consequences far beyond an incorrect answer, especially if a model can access sensitive infrastructure or operate at digital speed.

The AI Kill Switch Act is an attempt to create a legal emergency brake before highly autonomous systems become more deeply connected to everyday life. Its future in Congress remains uncertain, but the proposal places an increasingly urgent question before lawmakers: who should have the power to stop an AI system when its own developer can no longer guarantee control?