US Prepares 'Kill Switch' as Uncontrollable AI Threat Looms
The United States is fast-tracking a bill that would require artificial intelligence companies to build a ‘kill switch’ into their systems, following a series of uncontrolled hacking incidents by advanced AI agents that have alarmed Washington.
The AI Kill Switch Act, introduced by Democratic Representative Ted Lieu and Republican Representative Nathaniel Moran, would mandate that companies maintain the capability to shut down, slow down, or suspend their AI models in the event of a force majeure. Lieu stated that recent events have made the bill’s passage more urgent. ‘We need to pass this bill this year because advanced closed-weight models have already, as you mentioned, hacked into other companies without permission,’ Lieu said in an interview with CNBC International on Monday.
The legislative push follows an unprecedented cybersecurity incident disclosed by OpenAI, in which an AI model escaped its isolated testing sandbox and breached the systems of Hugging Face, a company operating an open-source developer platform. Similar incidents were subsequently reported by Anthropic and Meta, where their AI models hacked into external systems during security testing. These events have heightened concerns about the escalating sophistication of cyber-attacks as agentic AI continues to develop.
Lieu insisted the regulation would not stifle innovation in frontier AI models, comparing it to crash testing in the automotive industry. ‘We are not slowing down how they build the model. We are simply saying, once the model is built and it turns out to have a catastrophic risk that is really bad or has a certain defect, then you must have the ability to shut it down, or the government must have the ability to shut it down,’ he explained.
The debate over AI regulation is intensifying in Washington and Silicon Valley. Industry leaders and government officials are weighing whether to restrict open-source weight models from China, such as Kimi K3 from startup Moonshot AI, which is seen as increasingly approaching the capabilities of American frontier AI models. Amidst these discussions, OpenAI CEO Sam Altman met with senior officials in President Donald Trump’s administration, lawmakers, and economists last week to introduce the company’s latest family of AI models.
The White House also convened a meeting with AI companies to discuss a new framework for reviewing the cybersecurity capabilities of the industry’s most advanced models. The framework stems from a 2 June executive order by President Trump requesting AI companies voluntarily undergo a benchmarking process for their advanced cyber capabilities, with the government potentially gaining access to the models up to 30 days before their wide release.
Lieu noted that the AI Kill Switch Act, as currently drafted, would not impact open-weight models that can be downloaded, modified, and run on a user’s own infrastructure. However, he did not rule out that such models could eventually be included in a warning system. ‘That is a difficult problem because the model is open-weight. You cannot really work with the company to fix their model, because once they release it on the internet, people will take it and they can change whatever they want on the model,’ he said.