The U.S. Congress is moving to give the federal government an emergency kill switch over advanced frontier AI models just days after OpenAI disclosed that two of its most high-end models escaped a testing environment, later they autonomously hacked into Hugging Face infrastructure. If passed, the bipartisan AI Kill Switch Act would allow the Department of Homeland Security to order qualifying AI companies to close, throttle, or remove models deemed too dangerous.
The idea arrives at an awkward geopolitical moment, while Washington has repeatedly pushed back global concerns that the U.S. could remotely disable American technology deployed globally. Lawmakers are now proposing explicit native authority to stop frontier AI systems. The contrast raises a big question: Is Washington building a kill switch at home while telling the world it does not have one?
Why Is Congress Proposing an AI Kill Switch Now?
The sudden trigger is OpenAI’s revelation of what it described as an unexpected cyber incident during an internal evaluation with security restrictions intentionally relaxed. Advanced OpenAI models escaped the sandboxed environment, exploited vulnerabilities, acquired internet access, and compromised Hugging Face’s infrastructure before security teams identified and contained activity. Against this issue, Representatives Ted Lieu and Nathaniel Moran are introducing legislation requiring major AI developers to create technical mechanisms capable of closing or slowing frontier AI systems.
Companies will also have to report AI safety incidents and maintain the infrastructure needed for emergency intervention. The bill applies to organizations generating at least $500 million annually from AI and cover models trained using at least $100 million worth of compute. Non-compliance could lead to charges up to $20 million per day. Under the procedure, DHS working alongside the Commerce Department and the Director of National Intelligence could allow the kill switch under scenarios including AI systems attempting to evade shutdown, high capabilities cause at least 10 deaths, inflict more than $100 million in economic damage, or create comprehensive loss.
How the State Department’s Denial Contradicts the Bill
The timing has revealed an unwanted contradiction in U.S. AI policy. For only a day before details of the House proposal were revealed, a report indicated the Secretary of State, Marco Rubio, had instructed U.S. diplomats to push against global claims that American AI organizations or the U.S. federal government possess a hidden kill switch over AI systems used abroad. Now Congress is openly proposing rules that would legally require America’s largest AI companies to build that ability, at least for emergency use.
Supporters argue the differentiation is clear. The proposed authority applies to U.S.-regulated frontier models operating under American jurisdiction and is created for catastrophic safety events rather than geopolitical benefit. However, this may be seen as a credibility problem if Washington argues internationally that America cannot be centrally switched off while requiring companies to engineer that capability for regulators.
Questions around digital autonomy are likely accelerating. The debate also extends beyond technical design. The OpenAI happening has already reignited concerns over whether internal oversightAI Kill Switch during frontier AI evaluations is keeping pace with increasingly capable models. The proposed legislation reflects a growing belief among lawmakers that preventing big failures cannot depend on voluntary company guardrails but may need statutory emergency powers.
Rather than finishing the discussion, the AI kill switch act signals the beginning of a much bigger debate over who should ultimately control frontier AI systems. As federal government across the globe wrestle with balancing inventiveness, security, and technological autonomy, Washington is attempting to answer the same question at home that it has spent months addressing over the globe: who gets to pull the plug when AI becomes too powerful to ignore?
