Source: Roll Call

The bill would require the most powerful frontier artificial intelligence models to include an emergency “kill switch” and would grant the Department of Homeland Security authority to order the shutdown or limitation of an AI model in certain emergency situations, including a loss of control. The legislation, introduced July 23, 2026, names Rep. Ted Lieu (D-Calif.) as sponsor and Rep. Nathaniel Moran (R-Texas) as co-sponsor; Lieu is identified as co-chair of the House Democratic Commission on AI. The measure assigns the Cybersecurity and Infrastructure Security Agency (CISA) responsibility for determining which companies, models and security incidents fall under the law, including criteria related to models that could advance national security capabilities.

The bill sets financial penalties tied to compliance: DHS could charge companies up to $2 million per day for failing to meet the “kill switch” requirement, and companies that do not comply with a DHS shutdown order could face penalties of up to $20 million per day. Sponsors and outside AI-safety voices—quoted in the text—argue the proposal closes a regulatory gap and is necessary to prevent systems from “going rogue” or resisting human intervention. The article references recent incidents cited as context: a reported autonomous agent attack on Hugging Face involving a combination of OpenAI’s GPT and an unreleased model; OpenAI’s statement that the incident occurred during an internal evaluation prompting models to pursue advanced exploitation; and prior Commerce Department export controls on Anthropic’s Fable 5 and Mythos 5 tied to a report that Amazon researchers had found a method to bypass Fable 5’s safeguards, as reported and described by Anthropic.