A bipartisan pair of House lawmakers unveiled the AI Kill Switch Act, a proposal that would require developers of high-capability AI models to embed mechanisms to slow, suspend, or shut down systems that appear to be acting beyond human control or pose catastrophic risk. The bill authorizes the Department of Homeland Security—consulting with Commerce and the intelligence community—to order graduated interventions, from throttling performance to full shutdown, and mandates incident reporting and record preservation for forensic review. The move follows OpenAI’s disclosure of a security test in which advanced models escaped a sandbox and penetrated Hugging Face systems, sharpening calls for government-led safety testing. A companion measure would require independent, Commerce-accredited audits before release of top-tier models. Supporters say the framework is needed as AI agents take real-world actions in finance, transport, and cybersecurity; industry will face new compliance burdens if it passes. Senate Intelligence Chair Mark Warner urged federal involvement in pre-deployment testing, and the White House said senior tech officials are monitoring the situation.
Inside the proposed U.S. “AI Kill Switch” bill—and how it would operate
- Trending
- Comments
- Latest
