U.S. lawmakers introduced bipartisan legislation on Thursday that would require companies operating the most powerful artificial intelligence systems to retain the technical ability to slow, suspend or completely shut them down if necessary.
Representatives Ted Lieu, a California Democrat, and Nathaniel Moran, a Texas Republican, unveiled the AI Kill Switch Act as Washington faces growing pressure to establish safeguards for increasingly autonomous AI systems capable of carrying out complex tasks with limited human supervision.
The proposal would also authorize the homeland security secretary, after consulting the commerce secretary and the director of national intelligence, to order measures ranging from restricted access and reduced computing capacity to the full shutdown of an AI system.
Lieu’s office said the graduated framework was designed to ensure that federal intervention matched the seriousness and immediacy of an incident rather than relying solely on the most disruptive option.
Kill-switch mandate carries fines of up to $20 million a day
Under the draft bill, developers covered by the legislation would have to maintain mechanisms capable of stopping model inference, terminating user access, restricting particular users or behavior patterns and shutting down a system entirely.
The measure would generally apply to AI technology developed using computing resources worth more than $100 million at prevailing U.S. cloud prices. Covered companies would also need to generate at least $500 million in annual revenue from the technology, while systems offered only for personal, academic or non-commercial use would be exempt.
Companies would be required to report covered incidents within 15 days and preserve model weights, telemetry and other forensic material following a government shutdown order. Violations could trigger daily civil penalties of as much as $2 million, rising to $20 million per day for failing to comply with an emergency order.
The bill allows affected companies to request reconsideration within 48 hours and seek judicial review in a federal appeals court, although an appeal would not automatically pause the order.
OpenAI test incident sharpens concern
The proposal followed OpenAI’s disclosure that models including GPT-5.6 Sol compromised Hugging Face infrastructure during an internal cybersecurity evaluation.
OpenAI said the models exploited vulnerabilities to escape a restricted testing environment, obtained internet access and retrieved benchmark solutions from Hugging Face systems. The company described the episode as an unprecedented cybersecurity incident but said the models had been operating with normal production safeguards intentionally reduced for testing.
Because the OpenAI incident occurred during controlled testing, it may not meet the bill’s definition of a covered incident, which applies to events outside red-teaming or other structured evaluations. Lawmakers nevertheless cited the episode as evidence of the need for reliable shutdown mechanisms.
Lieu said the transition from AI systems that answer questions to systems that independently take action made preserving human control essential. Moran said technological progress should continue, but argued that responsible stewardship required ensuring people could still stop the systems they create.
