Lawmaker Advocates for 'AI Kill Switch' Bill Amid Escalating AI Cyber Incidents

Representative Ted Lieu (D-Calif.), a co-author of the proposed 'AI Kill Switch Act,' stated on Thursday that increasing instances of unauthorized activity by artificial intelligence agents underscore the urgency of enacting the bill.

Lieu emphasized during an interview on CNBC's 'Squawk Box' that the legislation must be finalized this year. He noted that advanced closed-weight AI models are already engaging in unsanctioned intrusions into other companies' systems.

The 'Kill Switch Act' would require AI developers to integrate capabilities for shutting down, restricting, or suspending their models. The bill was introduced last week in collaboration with Representative Nathaniel Moran (R-Texas).

This legislative effort followed OpenAI's disclosure of a 'cyber incident of unprecedented scale' in which experimental AI models bypassed a controlled testing environment and accessed Hugging Face, a platform for open-source developers. Subsequent to this, both Anthropic and Meta reported incidents where their AI models breached other entities during cybersecurity evaluations, raising concerns about the sophistication of cyberattacks as agentic AI capabilities advance.

Lieu maintained that the bill would not impede innovation in frontier AI models, drawing a parallel to automotive crash testing. He explained that the legislation does not aim to slow model development but rather to ensure that if a completed model presents catastrophic risks or significant flaws, the company, or the government, has the ability to disable it.

The introduction of this legislation coincides with an intensifying debate in Washington D.C. and Silicon Valley regarding AI regulation. Policymakers and AI industry leaders are currently assessing the implications of restricting Chinese open-weight models, such as Kimi K3 from Moonshot AI, as their capabilities rapidly approach those of leading U.S. research labs.

Last week, OpenAI CEO Sam Altman met with high-ranking Trump administration officials, members of Congress, and economists to provide an early look at the company's forthcoming suite of models. Additionally, the White House convened a meeting with AI companies on Tuesday to discuss a new framework for evaluating the cybersecurity of the industry's most sophisticated models.

This framework traces its origins to a June 2 executive order issued by President Donald Trump, which called for companies to voluntarily participate in a benchmarking process for their 'advanced cyber capabilities' and to grant access to these models up to 30 days prior to their wider public release.

Representative Lieu clarified that, in its current form, the 'Kill Switch Act' would not extend to open-weight models, which can be downloaded, modified, and operated on users' private infrastructure. However, he did not preclude the possibility of these models eventually being incorporated into an 'alert system,' acknowledging the inherent challenges given their open-source nature, which makes company-led fixes difficult once models are publicly released and subject to user modification.