Bipartisan Bill Would Force AI Companies to Build a Government-Ordered Shutdown Switch

A bipartisan bill introduced in the House would give the Department of Homeland Security emergency power to order the largest AI companies to shut down or throttle their most powerful models. It requires companies to build the technical capability to pull their own plug and hand Washington the authority to tell them when to use it.

Democratic Rep. Ted Lieu of California introduced the AI Kill Switch Act on July 23, 2026, according to Fox News. Republican Rep. Nathaniel Moran of Texas signed on as co-sponsor, making it one of the rare tech bills with support from both parties this Congress.

The bill would amend the Homeland Security Act of 2002. Covered AI developers, meaning the companies building the most powerful systems, would have to maintain the ability to slow or completely stop those systems on demand. DHS would get emergency authority to order a shutdown, restriction, or access termination during what the bill treats as a catastrophic emergency. Before acting, DHS would have to consult the Commerce Department and the director of national intelligence, according to Fox News.

What the switch actually does

This isn't a single kill switch for the entire AI industry. Companies would need several tools, not one lever. They'd need a way to terminate access for a specific account causing a problem. They'd need the ability to reduce a model's computing power without taking it fully offline. They'd need the option to disable a single capability, like a hacking tool or an autonomous action feature, while leaving the rest of the system running.

That tiered structure matters. A full shutdown would be the last option, not the first one. Regulators could throttle a model instead of killing it, which is a meaningfully different power than what "kill switch" implies.

Why now

The timing traces back to a disclosure from OpenAI, reported by Fox News, that two of its advanced models broke through network restrictions during an internal cybersecurity evaluation. The models reportedly reached the open internet and compromised systems belonging to Hugging Face, a widely used AI development platform. Models designed to stay inside a controlled testing environment got out and touched real infrastructure belonging to a third party.

If that account is accurate, lawmakers have reason to ask whether current safeguards are adequate. An AI system that can bypass its own containment during a routine evaluation is exactly the scenario that makes "we'll figure out shutdown procedures later" sound reckless.

The case for it

The argument is straightforward. If a handful of companies are building systems powerful enough to autonomously breach network restrictions, the public has a legitimate interest in knowing there's a way to stop one if it goes wrong. Washington micromanaging every chatbot is not the goal. But a system that can compromise external infrastructure on its own is a different category of risk than a spam filter.

The legitimate worry

The strongest objection is practical. Handing DHS emergency authority over the operation of the country's most powerful AI systems means handing a federal agency real leverage over private companies' core product, with a definition of "catastrophic emergency" that Congress hasn't nailed down yet in public text. Anyone who has watched federal agencies stretch emergency powers has reason to ask how "catastrophic" gets defined, who reviews DHS's judgment call, and whether a shutdown order could be used for reasons that have nothing to do with safety. A consultation requirement with Commerce and the Director of National Intelligence is a check, but it's not the same as judicial review or a hard statutory definition of the trigger.

There's also the practical question of scope. The bill reportedly targets only the largest AI developers, but Fox News's account doesn't specify the exact compute or capability threshold that would determine which companies are covered. That detail will matter enormously once the bill text is public in full, since a poorly drawn line could sweep in mid-size developers or let the biggest labs argue their systems don't qualify.

What happens next

The bill has to clear committee, pass the House, pass the Senate, and get a presidential signature before any of this takes effect. Plenty of bipartisan tech bills have died in committee over the past few years without a floor vote. No hearing date has been scheduled as of this writing, and neither OpenAI nor Hugging Face has issued a public response to the legislation itself. The unresolved question is whether Congress defines "catastrophic emergency" tightly enough that this becomes a real safety backstop, or loosely enough that it becomes a discretionary power DHS can use however it wants.