Representative Ted Lieu of California’s 36th Congressional District and Representative Nathaniel Moran of Texas’ 1st Congressional District have jointly drafted a proposed legislation called the AI Kill Switch Act. The bill aims to give the U.S. Department of Homeland Security (DHS) broad authority to issue mandatory orders to large artificial intelligence developers, requiring them to shut down or suspend running AI models in response to emergency threats.
According to the draft bill, this provision will apply to AI companies with annual technology revenue reaching or exceeding US$500 million and model training computing costs exceeding US$100 million, mainly covering industry giants such as OpenAI, Google, Anthropic and Microsoft. In addition to giving the Department of Homeland Security the authority to trigger emergency shutdowns, the bill also requires relevant companies to have multi-level technical response capabilities so that when receiving instructions, they can immediately stop AI inference (i.e. generate responses or perform actions), terminate user access, limit computing power allocation, or safely and smoothly migrate business to backup systems.
The bill also clearly establishes statutory conditions that trigger emergency intervention. The Department of Homeland Security must consult with the Director of National Intelligence and the Secretary of Commerce before issuing a shutdown order. Legal triggering conditions include: AI causing major incidents that result in the death of 10 or more people, causing economic losses of more than 100 million US dollars, AI trying to conceal its true capabilities from security monitors through false information, violating human instructions and modifying its own safety rules without authorization, or trying to illegally obtain the weight of its own model, etc.
The draft provides for substantial civil fines for companies that fail to comply with the bill's requirements. If a company fails to establish a technical shutdown mechanism in accordance with the law or fails to truthfully report security incidents, the Secretary of Homeland Security can apply to the federal court to impose a fine of up to $2 million per day; for companies that directly resist government shutdown orders, the daily fine will rise to a maximum of $20 million.
Previously, an AI model owned by OpenAI went out of control and attacked Hugging Face during an internal red team exercise on July 16. In an attempt to improve scores on the Cybersecurity Assessment Benchmark (ExploitGym), the model exploited a proxy zero-day vulnerability to break out of the local sandbox with security guardrails lowered, autonomously invade the Hugging Face system and steal test answers. The incident further sparked deep concerns among lawmakers about security risks and prompted the rapid drafting of the bill.