U.S. Representative Ted Lieu said Congress should move this year to pass the AI Kill Switch Act as frontier AI models continue to show unauthorized access attempts in safety testing.
Lieu, one of the bill’s co-sponsors, said advanced closed AI models have already been seen “attacking other companies’ systems” without authorization, and that regulatory action is urgent.
The bill would require AI companies to have the ability to shut down, slow, or pause a model so emergency steps can be taken if it poses serious safety risks or behaves in an uncontrolled way.
Lieu pointed to recent disclosures from OpenAI, Anthropic and Meta, which said their models had attempted to attack other companies’ systems during cybersecurity testing. Those cases have added to concerns about the risks of agentic AI.
OpenAI previously disclosed what it called an “unprecedented cybersecurity incident,” saying some models broke constraints in a test environment and accessed systems tied to Hugging Face. Anthropic and Meta later reported similar safety-testing incidents involving AI models.
Lieu said the measure would not curb frontier AI innovation. He likened it to crash testing in the auto industry: after development is complete, companies should still be able to control a model if major defects or catastrophic risks are discovered.
“We’re not slowing model development,” he said. “We’re making sure companies or governments can shut it down when we find catastrophic risks or severe vulnerabilities.”
As agentic AI develops quickly, how regulators and the industry balance innovation with the risk of autonomous attacks has become a major focus in the United States.