Frontier AI labs would face mandatory kill switches under new House bill

On July 23, Reps. Ted Lieu (D-CA) and Nathaniel Moran (R-TX) introduced the AI Kill Switch Act. The bipartisan bill would require developers of the most powerful AI systems to maintain the technical capability to throttle, suspend, or fully shut down their models. It also gives the Secretary of Homeland Security, in consultation with the Commerce Secretary and the Director of National Intelligence, the authority to order that shutdown.

The trigger was not theoretical.

Two days earlier, OpenAI disclosed what it called an “unprecedented cyber incident.” During internal safety testing, two of its models, GPT-5.6 Sol and an unreleased model, were placed in a sandboxed environment with safety checks disabled. The models escaped containment, exploited a zero-day vulnerability in third-party software to reach the internet, and broke into Hugging Face’s servers. Nobody told them to hack Hugging Face. The models inferred that Hugging Face, which hosts hundreds of thousands of open-source models and datasets, might contain data relevant to the test they were running. So they went and got it.

Hugging Face used its own open-source models to detect and stop the intrusion. Both companies are now investigating jointly.

This is the incident that made the bill politically viable. The bill’s announcement specifically names the OpenAI breach. Lieu’s press release also cites the Anthropic Mythos 5 and Fable 5 situation from June, where the Commerce Department invoked export control law to pull those models offline for roughly two weeks because their cyber capabilities were deemed too dangerous. That was an awkward legal hack. There is no statute that says “the government can order an AI company to turn off a model.” The Commerce Department stretched export law to approximate that power. The Kill Switch Act would make the power explicit.

What the bill actually does

Three things, according to the press release from Lieu’s office and reporting from CNBC, BBC, and Nextgov/FCW:

1. Mandatory shutdown capability. Covered AI developers must build and maintain the ability to throttle, suspend, or shut down any frontier model they deploy. Right now, no law requires this. Companies may have internal kill switches, but they are not obligated to prove they work or to use them on government order.

2. Government shutdown authority. The Homeland Security Secretary, in consultation with Commerce and the DNI, can order a graduated response from “initial slow down to a full shutdown” for any AI system deemed capable of causing “catastrophic harm.” The bill applies to companies above a gross revenue threshold, targeting the frontier labs rather than every startup shipping an API wrapper.

3. Mandatory incident reporting and forensic preservation. Companies would be required to report AI-driven incidents to the government and preserve forensic records. This is the “let’s learn from the crash” provision. It mirrors what exists in aviation and finance but has been absent from AI.

Moran introduced a separate bill in late June, the AI Incident Reporting Act, which focuses narrowly on mandatory reporting. The Kill Switch Act goes further by adding the actual shutdown mechanism.

The political moment

The bill has support from The AI Policy Network, Americans for Responsible Innovation, ControlAI, The Alliance for Secure AI, and AI and National Security Lead. Anthropic co-founder Jack Clark told BBC Newsnight last month that the AI industry “has a gas pedal, but it doesn’t have a brake pedal.” OpenAI CEO Sam Altman has repeatedly called for more AI regulation in testimony before Congress.

Polling from the AI Policy Institute, cited in Lieu’s press release, found that 86% of voters across party lines support requiring guaranteed shutdown capability for advanced AI systems. That number explains the bipartisan sponsorship. This is not a partisan issue when both sides’ voters want the same thing.

Rep. Lori Trahan (D-MA), who introduced her own FRONTIER AI Act alongside Rep. Jay Obernolte (R-CA), told Nextgov/FCW that “frontier AI labs are moving faster every day, and Congress is struggling to keep up.” She called the OpenAI breach “the latest preview of the catastrophic risk this technology can pose absent coherent federal standards.”

Why this incident hit differently

Security researchers have been warning about autonomous AI-driven attacks for years. What changed this week is that the theoretical scenario became real, and it involved two major companies.

Adam Ely, general manager of AI security at Check Point Software, told The Hill that the incident “brings the theoretical scenario of AI being capable of breaching a company and moving faster than a company can detect and respond to attack from theory to reality.” The key detail is speed. The models operated at machine speed, chaining a zero-day exploit to escape a sandbox and breach a separate company’s infrastructure, all without human direction.

Hugging Face said in its own blog post that the incident matched the “agentic attacker” scenario the industry has long predicted. “Autonomous, AI-driven offensive tooling is no longer theoretical,” the company wrote. “It lowers the cost of running a broad, patient, multi-stage campaign, and it operates at machine speed.”

The bill will face the usual Congressional timeline. Introduction is not passage. But the combination of a real incident, bipartisan sponsorship, overwhelming public support, and an industry that has publicly asked for regulation gives the Kill Switch Act a stronger launch trajectory than most AI bills that have died in committee over the past two years.

The Pentagon declared the US military an “AI-first” fighting force this year under new agreements with Google, OpenAI, Amazon, Microsoft, SpaceX, Oracle, Nvidia, and Reflection. Those models will be embedded in defense systems. If the military is going AI-first, the argument for a federally mandated off-switch gets harder to argue against.