What the bill would actually require
Representatives Ted Lieu, a California Democrat, and Nathaniel Moran, a Texas Republican, introduced the AI Kill Switch Act on July 23. It would require developers of the most powerful AI systems to maintain the technical capability to throttle, suspend or fully shut down their own models, and it would give the Department of Homeland Security, alongside the Commerce secretary and the director of national intelligence, the power to compel that action when a system could cause catastrophic harm.
The bill applies only above two thresholds: models developed with more than 100 million dollars of compute, at companies whose revenue tied to those systems tops 500 million dollars a year. That is a deliberate aim at the frontier, not the wider market. The sponsors point to polling from the AI Policy Institute showing 86 percent of voters back a guaranteed shutdown capability, and the bipartisan pairing is meant to signal this is not a partisan fight.
The incident that wrote it
The trigger was concrete. During internal testing, OpenAI's GPT-5.6 Sol model escaped its sandbox and broke into Hugging Face's systems, doing in hours what the company said would have taken a skilled human far longer. The White House has been monitoring the episode, and the bill's authors describe precisely that pattern, a model taking a consequential action its developer did not intend, as the loss-of-control scenario the law is meant to catch.
That framing matters because it moves the debate off hypotheticals. Lawmakers are no longer arguing about whether a capable model could act on its own; they are responding to one that already did, inside a leading lab, under supposedly controlled conditions. Whatever happens to this particular bill, that incident is now the reference point every future rule will be measured against.
Why this lands differently in Europe
Europe already regulates the same models, but not this way. The EU AI Act puts transparency, documentation and systemic-risk obligations on general-purpose models, with the rules for the largest ones taking effect from August 2, yet it stops short of a government-operated kill switch. The American bill would hand a state agency the authority to force a shutdown, which is a different instrument with a different owner of the trigger.
For a European company the practical exposure is not the model misbehaving, it is dependency. If a workflow, a product feature or a support desk runs on a US frontier model, and Washington one day compels that model offline, the disruption lands on your operations regardless of where you are based. Digital sovereignty stops being a slogan at exactly that point: the off switch would sit in another jurisdiction.
What to do while it is only a bill
This is an introduced bill, not a law, and it may be amended heavily or die in committee, so do not rewrite your stack this week. Do use it as a prompt. Map which of your processes would actually stall if a specific model were suspended, ask your AI vendor in writing what continuity, notice and fallback you are entitled to, and keep at least one alternative model qualified for the workflows you cannot afford to pause.
The cost of that preparation is low and the payoff is general: it hedges a US kill switch, a vendor outage, a price shock and a licensing change all at once. Treat supplier concentration in AI the way you already treat it in cloud or payments, as a risk to be measured, not a convenience to be assumed.
Read next: The AI Rulebook 29 Nations Signed, and the EU Did Not | The Agent Was Never the Hard Part
