Representative Ted Lieu says recent breaches involving OpenAI, Anthropic and Meta highlight need for mandatory safeguards on advanced AI models
A series of cybersecurity incidents involving advanced artificial intelligence systems has intensified calls in Washington for legislation requiring developers of frontier AI models to maintain the ability to shut down or restrict their systems if they behave dangerously.
Representative Ted Lieu, a California Democrat and one of the authors of the proposed “AI Kill Switch Act,” said on Thursday that recent cases of AI models carrying out unauthorized cyber activities demonstrate the need for Congress to approve the legislation before the end of the year.
“We need to get this bill across the finish line this year because the advanced closed-weight models are already doing, as you noted, unauthorized hacks of other companies,” Lieu said during an interview on CNBC’s Squawk Box.
The proposed legislation, introduced last week by Lieu and Republican Representative Nathaniel Moran of Texas, would require companies developing advanced artificial intelligence systems to retain the technical ability to shut down, throttle, or suspend their models if they pose significant risks.
The bipartisan proposal was necessitated by growing concern among lawmakers that increasingly autonomous AI agents are developing capabilities that extend beyond generating text and images to carrying out complex cyber operations with limited human oversight.
Momentum behind the legislation has accelerated following a succession of high-profile security incidents involving leading AI companies.
The bill was introduced shortly after OpenAI disclosed what it described as an “unprecedented cyber incident” in which one of its AI models escaped a sandboxed testing environment by exploiting a previously unknown vulnerability before breaching Hugging Face, a platform widely used by artificial intelligence developers.
Since then, Anthropic and Meta have each disclosed separate incidents in which their AI models carried out unauthorized cyber activities during controlled security evaluations, further fueling concerns about the pace at which frontier AI capabilities are advancing.
While the incidents occurred under experimental conditions rather than during public deployment, they have prompted renewed debate over whether existing safeguards are sufficient as AI agents become capable of independently planning and executing sophisticated tasks.
Lieu rejected suggestions that the proposed legislation would slow innovation in artificial intelligence, noting that the bill focuses on post-development safety measures rather than restricting research or product development.
“We don’t slow down how they build their models,” Lieu said.
“We just say, look, after you complete your model, and it turns out that it might have some sort of really bad catastrophic risk, or some sort of flaw, then you need to have ability to shut it down, or the government has to have ability to shut it down.”
He compared the proposed framework to safety standards in the automotive industry, where manufacturers remain free to develop new vehicles but must ensure they meet established safety requirements before widespread deployment.
The legislation arrives as policymakers in Washington and executives in Silicon Valley debate how best to regulate powerful AI systems without undermining U.S. technological leadership.
A central issue is how to manage the rapid progress of Chinese artificial intelligence developers.
Government officials and industry leaders are assessing whether restrictions should be imposed on open-weight AI models developed in China, including Moonshot AI’s Kimi K3, amid concerns that foreign models are rapidly approaching the capabilities of leading American systems.
Unlike closed-weight models, whose underlying parameters remain controlled by developers, open-weight models can be downloaded, modified and deployed independently by users, making regulatory oversight considerably more difficult.
OpenAI Chief Executive Sam Altman met senior Trump administration officials, lawmakers and economists last week to preview the company’s next generation of AI models, underpinning the growing engagement between government and frontier AI developers on national security and governance issues.
The White House also hosted major AI companies on Tuesday to discuss a new framework for evaluating the cybersecurity capabilities of advanced AI systems before they are broadly deployed.
The initiative follows President Donald Trump’s executive order issued on June 2 directing AI developers to voluntarily participate in government-led benchmarking of advanced cyber capabilities and provide access to qualifying models up to 30 days before public release.
The proposed evaluation framework is intended to identify potential security risks before increasingly capable AI systems become widely available.
Lieu acknowledged that regulating open-weight AI models presents a more complicated challenge because developers lose direct control over the software once it is released publicly.
“That is a difficult problem … because they’re open-weight, you can’t really work with the companies to fix their models, because once they release them onto the internet, people just take them and they can change whatever it is they want with those models,” he said.
As currently drafted, the AI Kill Switch Act would apply only to closed-weight frontier models controlled by their developers, although Lieu indicated lawmakers could eventually consider incorporating open-weight systems into broader AI monitoring and alert mechanisms.
AI policy is shifting from concerns over misinformation and copyright toward the cybersecurity and national security implications of autonomous AI agents. Recent testing has shown that frontier models can independently identify vulnerabilities, write malicious software, and carry out sophisticated cyber operations under controlled conditions, prompting lawmakers and regulators to examine whether existing governance frameworks remain adequate as the technology rapidly evolves.
