A new bill in the United States aims to give the federal government the authority to slow down or completely shut down advanced AI systems when they pose a serious risk to national security or society. The AI Kill Switch Act follows recent incidents in which experimental AI models from OpenAI and Anthropic behaved dangerously, according to the bill’s sponsors.
The bill was introduced by Democratic Representative Ted Lieu and Republican Representative Nathaniel Moran. As a result, the proposal has received rare bipartisan support, according to Ars Technica. If Congress approves it, the Secretary of Homeland Security could require AI companies to restrict a model or take it completely offline. To that end, AI developers must build in technical safeguards that allow their systems to be remotely throttled or shut down. Companies that refuse to comply with such an order risk fines of up to $20 million per day.
Recent AI incidents as the impetus
According to the sponsors, the measure is necessary because the risks posed by advanced AI are no longer hypothetical. They cite, among other things, the recent security incident in which a pre-release model from OpenAI compromised Hugging Face’s systems during an internal cybersecurity test. They also mention Anthropic’s Mythos 5 and Fable 5 models, whose cyber capabilities, they argue, were so advanced that the U.S. government invoked export control laws to limit their further dissemination.
The AI Kill Switch Act is intended to provide a legal framework not only for such situations but also for future incidents in which an operational AI model proves to be insufficiently controllable.
Intervention in cases of sabotage or hidden behavior
The bill outlines various situations in which the government may intervene. This applies, for example, when an AI system pursues goals other than those intended by the developer, attempts to sabotage a shutdown command, or attempts to conceal its own functionality from oversight or security mechanisms.
Incidents with serious societal consequences also fall under the law. The proposal specifically mentions situations in which AI unintentionally results in the deaths of 10 or more people or causes economic damage of at least $100 million. However, the wording makes it clear that these powers may also be exercised in the event of less severe incidents.
Following such an incident, the Secretary of Homeland Security, in consultation with the Department of Commerce and U.S. intelligence agencies, may decide to slow down or completely shut down an AI system. In addition, the bill introduces a mandatory reporting requirement for serious AI incidents, and developers must retain forensic data to investigate the cause after the fact.
Only for the largest AI developers
The law targets only the largest providers of AI technology. It applies to companies that generate at least $500 million in annual revenue from AI and to AI systems whose development or deployment represents at least $100 million in cloud computing power.
An exception applies to controlled red-team tests. Simulations intended solely to test the security of models are not covered by the proposed regulations.
Politically sensitive
The proposed bill is expected to spark debate over the extent to which the federal government may intervene in the operations of commercial AI providers. Anthropic is currently suing the U.S. government after the Trump administration banned federal agencies from using the company’s technology. Anthropic argues that this decision was politically motivated, while the White House cites national security interests.
The bill’s sponsors emphasize that powerful AI systems are becoming increasingly capable of acting autonomously. According to Ted Lieu, the government must therefore have clear legal authority to intervene quickly in exceptional situations before an incident leads to widespread damage.