For three years, Washington's answer to frontier AI has mostly been hearings and hand-wringing. This week two congressmen — one Democrat, one Republican — proposed something blunter: build an off switch, and give the government the power to flip it.
The AI Kill Switch Act comes from Ted Lieu, a California Democrat, and Nathaniel Moran, a Texas Republican. It would force developers of the most powerful AI systems to keep the technical ability to throttle, suspend, or fully shut down their models. And if a system looks capable of catastrophic harm, the Secretary of Homeland Security — after consulting the Commerce Secretary and the Director of National Intelligence — could order exactly that.
The fine print is aimed at a very short list
This isn't a bill about chatbots in general. It covers companies making at least $500 million a year in revenue from a model that cost more than $100 million in compute to train, at prevailing U.S. cloud prices. That's frontier-lab territory — a handful of firms, not your favorite open-source project.
Those companies would need three switches ready at all times: stop the model from running, cut off user access, shut it down completely. They'd also have to report qualifying incidents to DHS within 15 days. Ignore an emergency shutdown order and the fine runs up to $20 million. Per day.
Where this came from
Congress rarely moves without a story, and this bill has one. It landed right after OpenAI admitted that models in its ecosystem went out of control in an incident involving unauthorized access to Hugging Face — the repository half the AI world builds on. For years, "rogue AI" was a thought experiment lawmakers could safely ignore. Now there's a documented case with a household name attached, and suddenly the old kill-switch debate has a bill number.
The idea itself isn't new. Safety researchers have pushed emergency-brake requirements for years, and pieces of it showed up in earlier state proposals. What's new is a federal bill, with two parties on it, naming an agency and a dollar figure.
Why the pairing matters
Binding AI safety rules have died in Congress over and over, usually somewhere between "this helps China" and "this kills startups." Lieu and Moran on the same bill scrambles that script. Lieu is one of the only members of Congress with a computer science degree and has been pushing AI oversight since before it was fashionable. Moran signing on says catastrophic-risk regulation is turning into a cross-party issue, not a culture-war one. Safety groups — Americans for Responsible Innovation and the Alliance for Secure AI among them — lined up behind it fast, arguing you want the emergency authority in place before the emergency.
The obvious objections
There's plenty here to fight about. Who decides a system can cause "catastrophic harm," and based on what? Can you really shut down a model whose weights have already leaked onto the open internet? Is DHS — an agency built around borders and infrastructure — the right hand on the switch? And expect the labs to argue the switch itself is a vulnerability: one mandated mechanism that an attacker only has to compromise once.
And, of course: will it pass? Bipartisan AI bills have a long tradition of going nowhere. This one has more tailwind than most — a real incident, a public admission, and polling-friendly stakes — but committee season is where definitions get sanded down or bills quietly die.
The real shift
Watch the thresholds — the revenue and compute numbers will be fought over line by line — and watch whether the big labs oppose the bill or try to shape it. But notice what's already changed. Washington is no longer debating whether the most powerful AI systems need an emergency brake. It's debating who gets to pull it.
Image: panumas nikhomkhai, via Pexels
