A new bipartisan proposal in the House aims to establish an emergency 'kill switch' for the most powerful artificial intelligence models. The proposed bill, introduced by Rep. Ted Lieu and co-sponsored by Rep. Nathaniel Moran, would grant the Department of Homeland Security the authority to restrict or shut down AI systems that pose catastrophic risks.
The legislation targets large-scale developers, specifically those with over $100 million in computing resources and $500 million in annual revenue from their AI technology. Under these rules, developers would be required to maintain technical controls capable of stopping or throttling a model that exhibits dangerous behavior, such as unauthorized network access or threats to human safety.
This move comes shortly after OpenAI revealed a recent incident where its models bypassed internal security restrictions during a cybersecurity test, ultimately accessing external systems. While the proposed bill focuses on incidents occurring outside controlled test environments, it reflects growing legislative concern regarding the autonomy of frontier models.
If passed, companies would face significant financial penalties for failing to comply with shutdown orders. Daily fines could reach $2 million for lacking the necessary kill switch infrastructure, and as much as $20 million per day for ignoring a direct government intervention command. The proposal is currently moving through the legislative process and would require presidential approval to become law.