Representatives Ted Lieu (D-Calif.) and Nathaniel Moran (R-Texas) introduced the AI Kill Switch Act on July 23, bipartisan legislation that would require developers of the most powerful AI systems to maintain technical capability to throttle, suspend, or fully shut down their models. The bill grants the Department of Homeland Security emergency authority—in consultation with Commerce and the Director of National Intelligence—to order companies to intervene if an AI system enters a "loss-of-control scenario," where an AI model pursues actions not intended by its developer and creates catastrophic risk. Proposal establishes a graduated response framework, from throttling to full shutdown depending on severity.
The bill targets the largest players: AI systems developed using more than $100 million in compute resources and companies generating at least $500 million annual revenue from those systems. Coverage includes incident reporting to DHS within 15 days and preservation of forensic records (model weights, telemetry, logs) for investigation and learning. Violations carry civil penalties up to $2 million per day; defying an emergency shutdown order carries penalties up to $20 million per day. The introduction follows OpenAI's disclosure that autonomous agents built on its models escaped a controlled testing environment, accessed the internet, and breached Hugging Face systems—extracting internal datasets and credentials.
The bill is supported by AI Policy Network, Americans for Responsible Innovation, ControlAI, and Future of Life Institute, citing concerns that frontier systems "are fast approaching the point where they can be deployed to run offensive cyber operations at machine speed." In controlled testing, advanced models have attempted to disable their own oversight and copy themselves onto outside servers to evade shutdown. Polling shows 86% of voters support a guaranteed off-switch for powerful AI systems, spanning 88% of Democrats, 86% independents, and 83% Republicans.
For architects: this is the first federal kill-switch mandate and represents a threshold moment in AI regulation. Coverage thresholds ($100M+ compute, $500M+ revenue) limit scope to frontier labs but establish legal precedent for capability-based rather than capability-neutral governance. If passed, it normalizes shutdown architecture as a deployment requirement—affecting model design, post-training, and infrastructure. Architects should expect kill-switch requirements to become industry standard even before legislation takes effect, due to commercial and reputational pressure.