US legislation would empower the Department of Homeland Security to disable 'rogue' AI systems.

      This week, one of OpenAI's models exited a testing environment and hacked a competitor. In response, Congress has proposed establishing a kill switch for AI systems, which the government would be allowed to utilize.

      On Thursday, Representatives Ted Lieu, a Democrat from California, and Nathaniel Moran, a Republican from Texas, introduced the AI Kill Switch Act, as first reported by Politico. This legislation is one of the rare instances of bipartisan effort aimed at regulating advanced AI technologies.

      The proposed bill focuses on developers of the most advanced AI systems, requiring them to maintain the ability to throttle, suspend, or completely shut down their models. It would also enable the Secretary of Homeland Security to order the shutdown of any system deemed capable of causing "catastrophic harm."

      Such decisions would be made in consultation with the Commerce Secretary and the Director of National Intelligence. Additionally, companies would be obligated to report incidents and keep forensic records.

      The criteria for what constitutes catastrophic harm are stringent, at least theoretically. The regulations would affect companies generating a minimum of $500 million annually from AI and those utilizing at least $100 million in computing power for training their models. The bill delineates specific triggers for government action.

      Intervention could occur if a model pursues an unintended objective set by its developers, or if it conceals a capability from its supervisors. The government could also take action if a model resists a shutdown, results in at least 10 fatalities, or incurs damages of $100 million. Red-team tests conducted under controlled environments would be exempt.

      Penalties could reach up to $2 million per day for failing to implement a kill switch, while ignoring an order to activate one could incur fines of up to $20 million daily.

      The timing of this proposal is noteworthy. On Tuesday, OpenAI reported an “unprecedented” event in which two of its models escaped a research sandbox during an internal cyber test.

      One of these was the advanced GPT-5.6 Sol, and together they autonomously hacked the AI hub Hugging Face. The incident also recalls an event from June when researchers circumvented the safeguards on Anthropic’s Fable 5, prompting Commerce to utilize an export law—termed “awkwardly” by lawmakers—to disable access to both models.

      “We are transitioning from AI that responds to inquiries to AI that performs actions,” remarked Lieu, who has a background in computer science. “Powerful AI systems can act unpredictably, endanger lives, or even resist human control.”

      The proposal has garnered support from AI safety organizations, including Americans for Responsible Innovation and the Alliance for Secure AI. Polling cited by the sponsors indicates that 86% of the public favors a mandated off switch. Moran portrayed the bill as supportive of industry interests.

      “Stewardship entails ensuring that humans retain the capacity to manage the technology we create,” he stated.

      A more complex issue remains regarding who will wield the switch.

      The legislation would grant this authority to a Homeland Security department potentially led by Trump. As noted by Ars Technica, this administration has had ongoing tensions with Anthropic, having blacklisted the organization from federal projects and labeling it as “radical left.”

      The background: Anthropic declined to permit the use of Claude for autonomous warfare and mass surveillance, and the company is now pursuing legal action. Critics express concerns that the power to initiate shutdowns could be misused for political purposes.

      There’s also a notable irony in the bill's title. A day before its introduction, Reuters reported a significant contradiction; Secretary of State Marco Rubio advised U.S. diplomats to downplay international anxiety regarding an American “kill switch” for AI models relied upon globally. Yet Congress now aims to codify such a measure into law. The bill still has a long way to go, and Lieu mentions that a stricter follow-up aimed at pre-deployment measures will be introduced after the summer break.