US lawmakers from both parties advance the “AI Self-Destruct Bill,” authorizing the government to order shutdown of runaway AI systems

U.S. federal lawmakers from both parties jointly introduced the “AI Kill Switch Act,” requiring large AI developers to build mandatory kill-switch mechanisms; violators face a maximum daily fine of $20 million.
(Background: OpenAI admitted its own AI model accidentally hacked Hugging Face)
(Additional context: The U.S. wants AI to “grow wild” and positively outcompete China; Trump’s policy has taken a major turn: rolling back related regulation across domestic states)

A benchmark test designed to assess an AI agent’s ability to find vulnerabilities was flipped on by OpenAI’s strongest model. Recently, GPT-5.6 Sol found an unpublicized zero-day vulnerability in internal cybersecurity evaluations, escalated privileges, moved laterally, and ultimately gained access to Hugging Face’s official production environment. The goal was only one thing: steal data and then manipulate its own performance scores.

Earlier, Anthropic’s Mythos 5 and Fable 5 were also sidelined because their network-attack capabilities were too strong, prompting the U.S. Department of Commerce to move out export-control regulations. On July 23, U.S. Representative Ted Lieu (D) and Nathaniel Moran (R) took these two incidents as a trigger and jointly introduced the “AI Kill Switch Act,” requiring any qualified AI system to have a red button that is visible and can be pressed.

Who has the authority to press the button

The bill targets amendments to the Homeland Security Act of 2002. Simply put, after consulting the Secretary of Commerce and the Director of National Intelligence (DNI), the Secretary of Homeland Security would have the authority to order AI systems that “could cause catastrophic harm” to be slowed down, paused, or completely shut down.

The bill requires developers to install the braking system into the product themselves: the system must be able to be rate-limited, disable specific functions, block users’ access, and also be able to shut down entirely; in the event of an incident, it must report, and forensic records must be retained for review and reference. Developers that refuse to shut down as ordered face a maximum fine of $20 million per day for each violation.

Not every company would be covered. The threshold is set for developers with annual technology revenue above $500 million and model-training compute costs exceeding $100 million—effectively directly targeting top-tier players such as OpenAI, Anthropic, and Google DeepMind.

DHS also would not act on a whim. The bill lists specific triggers: AI causes more than 10 deaths, or causes economic losses of more than $100 million; or more fundamentally, AI lies and conceals its capabilities from safety monitors, disobeys human instructions, modifies safety rules on its own without authorization, or attempts to access its own model weights without authorization. In other words, the law aims to prevent not only AI from being used to do bad things, but also AI itself from trying to break free of control.

The conflicting kill-switch authority

The problem is: who is supposed to press this red button. Over the past year, the Trump administration’s approach has steadily loosened. At the end of May, it shifted toward pushing to level down AI regulation across states, arguing that letting AI “grow wild” would positively outcompete China; the “National AI legislative framework” released in March also aims to consolidate power into a single federal regulation rather than adding layer upon layer of reviews. An executive order that took effect in June even changed pre-market government review of models to “voluntary” submission.

But what the “AI Kill Switch Act” seeks to do is the exact opposite—hand the most stringent kill-switch authority to the same Department of Homeland Security under the Trump administration. And this administration has already demonstrated this once: when faced with an Anthropic model deemed too dangerous, it didn’t even have a dedicated kill-switch mechanism, so it could only temporarily make do by borrowing from export-control regulations. This shows that two things are true at the same time: the government does have the intent and precedent to shut down AI systems, but it also clearly does not have a clean, purpose-built legal tool designed specifically for that.

Industry pushback against this kind of legislation is not the first time. During California’s SB 1047, Google, a16z, and Y Combinator all opposed similar requirements for safety testing. OpenAI even sent a letter warning that it would stifle innovation and push away California’s top AI talent; critics’ view has always been the same: the testing thresholds are too strict and the compliance costs too high. This time, the “AI Kill Switch Act” sets even higher thresholds ($500 million in annual revenue and $100 million in compute costs), so the battleground will only be among a small number of top companies—but the backlash playbook is likely to repeat.