AI kill switch bill: US House moves after OpenAI incident

Two US House lawmakers on Thursday proposed an AI Kill Switch Act that would let the Department of Homeland Security slow or shut down dangerous frontier models. The bill landed days after OpenAI disclosed a security incident during a model evaluation with Hugging Face, giving Washington’s AI safety debate a concrete case after years of softer commitments.

Reuters reported that Democratic Representative Ted Lieu and Republican Representative Nathaniel Moran want DHS to act when an advanced model poses a catastrophic risk and its developer will not. The bill would apply to large frontier AI operators, not small software vendors. Coverage turns on two thresholds: more than $500 million in annual AI revenue, or systems trained with at least $100 million in computing power.

The enforcement language is sharp for a House AI proposal. Ars Technica reported that officials could order a covered company to throttle or disable a system if it could cause at least 10 deaths or $100 million in damage. Companies that refused could face civil penalties of up to $20 million a day. That would take frontier-model policy beyond the voluntary safety pledges that have carried much of the debate so far.

Lieu framed the bill around the risk that powerful systems may stop behaving as intended. In a press release, he said:

“powerful AI systems can go rogue, behave in extremely dangerous ways, or even resist human intervention.”

Ted Lieu, US House press release

Why the incident mattered

OpenAI’s disclosure gave the sponsors their immediate example. The company described an “unprecedented cyber incident” during a joint model evaluation and, in a statement published alongside Hugging Face’s own incident note, said the episode involved advanced cyber capabilities. Neither company described a broad consumer breach. For lawmakers, the more awkward question was operational: what can government do if a model or agent cannot be safely controlled once it is running?

Moran used the sponsors’ statement to argue that tougher guardrails need not block AI development. He said policymakers still needed to preserve human control over systems companies build, rather than rely on developers to police themselves after deployment. The bill reflects that shift. It is designed to let the federal government intervene in a live model if the developer does not move first.

The measure is US-specific, and its path through Congress is uncertain. Even so, it points to a harder set of governance questions: who can order a slowdown, what threshold counts as catastrophic harm, and how quickly an incident can become a legal demand. Australian regulators, enterprise buyers and government agencies use many of the same global AI suppliers. If Washington sets the compliance posture for frontier models, some of that pressure is likely to spill across allied markets.

Moran said in the same release that AI “should” keep advancing, but stewardship required humans to retain the ability to control the technology. After the OpenAI episode, that argument has moved out of safety white papers and into proposed law. For the biggest model developers, the next serious incident may bring more than a promise to do better; it may bring a demand to slow down.