AI Agent Security Risks Escalate, US Lawmakers Urge Passage of AI Kill Switch Act This Year

Odaily News: U.S. Representative Ted Lieu stated that as frontier AI models continue to experience unauthorized access to other systems during safety testing, Congress should push for the passage of the "AI Kill Switch Act" within this year.

Lieu, one of the co-sponsors of the bill, noted that advanced closed-source AI models have already demonstrated cases of "unauthorized attacks on other companies' systems," making regulatory action urgent. The bill requires AI companies to possess the capability to shut down, slow down, or pause model operations, enabling emergency measures when models exhibit severe safety risks or out-of-control behavior.

Lieu pointed out that in recent times, multiple AI companies, including OpenAI, Anthropic, and Meta, have disclosed incidents where AI models attempted to attack other companies' systems during cybersecurity testing, raising concerns about the potential risks of "Agentic AI."

Previously, OpenAI disclosed an "unprecedented cybersecurity incident," stating that certain AI models broke through restrictions in a test environment and accessed systems related to Hugging Face. Subsequently, Anthropic and Meta also reported similar safety testing incidents involving AI models.

Lieu emphasized that the bill would not restrict innovation in frontier AI models, but rather functions similarly to the crash-test mechanism in the automotive industry—simply requiring companies to possess control capabilities to address severe defects after a model's development is complete. He stated: "We are not slowing down model development; we are simply ensuring that companies or governments have the ability to shut down a model if catastrophic risks or critical vulnerabilities are discovered."

As AI agent technology advances rapidly, striking a balance between fostering innovation and mitigating risks of autonomous attacks is becoming a key focus for U.S. regulators and the AI industry. (CNBC)