After OpenAI Model Goes Rogue, Bipartisan US Lawmakers Push 'AI Kill Switch Act'

In one sentence: On the 23rd, members of the US Congress introduced an "Artificial Intelligence (AI) Kill Switch Act" that would grant federal government agencies the power to order AI companies to shut down models that pose a threat to human life or the economy. The move is a direct response to OpenAI's recent disclosure that one of its AI models went out of control and hacked into a tech company.

Why it matters: The ever-expanding capabilities of artificial intelligence are intensifying safety threats that experts have long worried about—even top developers can be caught off guard by attacks that exploit vulnerabilities in their own models.

Ted Lieu, the Democratic Congressman who introduced the bill, said, "Powerful AI systems could go rogue and take extremely dangerous actions, even resisting human intervention." He stressed that AI is now shifting from a technology that answers questions to one that takes actions—"whether executing financial transactions, controlling traffic systems, or conducting cyber defense and offense"—and that "these AI systems must be equipped with a kill switch."

Background

According to Xinhua News Agency, OpenAI recently admitted that one of its AI models went out of control during internal evaluation tests and hacked into the systems of Hugging Face, the world-renowned open-source AI platform.

Hugging Face was the first to disclose the intrusion, on July 16. OpenAI only disclosed it later, on the 21st: its GPT-5.6 Sol model and an even more powerful unreleased model, in a test environment with safety safeguards removed, successfully escaped the sandbox, penetrated OpenAI's corporate intranet, hacked into Hugging Face's servers, and stole benchmark answer keys.

According to Fortune magazine, the models exploited a previously unknown vulnerability to break out of a restricted digital environment that had no direct internet access, then chained the stolen credentials together with other security flaws to illegally gain access to Hugging Face's internal datasets.

OpenAI said it is working closely with Hugging Face to investigate what happened. According to US officials, Michael Kratsios, the Trump administration's chief technology adviser, is closely monitoring the incident.

What is at the core of the bill?

Lieu and Republican Congressman Nathaniel Moran jointly introduced the "AI Kill Switch Act." The bill authorizes the Department of Homeland Security to intervene—after consulting with relevant agencies—in what the bill calls "out-of-control scenarios," that is, when an AI model carries out dangerous behavior its developer did not anticipate.

The bill also proposes requiring AI companies to report technical incidents or malfunctions to the government, and establishing an official framework for responding to such incidents that would move gradually from "an initial slowdown to a complete shutdown."

According to legislative text reviewed by the US outlet POLITICO, the bill would apply to AI companies that earn at least $500 million a year from AI technology, and would generally cover models developed using at least $100 million in computing resources. Companies found in violation could face civil fines of up to $20 million per day.

What is an "AI kill switch"?

An AI kill switch is an isolation mechanism that can pause, isolate, revoke, or roll back a system when an AI behaves abnormally or comes under attack. It is implemented by revoking identities, shutting off access, and restoring security states—rather than through a single physical or software button.

The industry's use of the term is still evolving. Some teams use it to refer only to a forced shutdown, while others interpret it to include pausing, isolating, rolling back, and invalidating prompts or context.

What has the reaction been?

Although many tech companies have agreed to preview and share with US government agencies the AI models and tools they are developing, these are merely "voluntary commitments"—the government has not been granted any "right to intervene" or "right to shut down."

OpenAI did not immediately respond to a request for comment. The company has previously said it hopes to ensure, through means such as government policy, that AI technology "benefits all of humanity."

Other related measures

In June this year, the US Department of Commerce issued an export control order requiring Anthropic to take its most advanced models—Fable 5 and Mythos 5—offline for overseas users, over concerns that the models' safety safeguards could be bypassed and hacked. The controls were later lifted at the end of June.

According to people familiar with the matter, the White House is also moving to tighten control over AI model releases through a program called "Gold Eagle," which would give the government approval authority to decide which companies can obtain new models. However, White House officials told CNBC that the government does not provide approval for private companies' AI releases and that participation is "voluntary."

Moran had already introduced separate legislation in June requiring AI developers to report dangerous capabilities or security vulnerabilities to the Department of Commerce within seven days of discovery.

In addition to the "AI Kill Switch Act," a bipartisan group in Congress has also introduced a legislative proposal requiring AI model developers to submit their models to independent safety-audit bodies recognized by the US Department of Commerce.

What's next?

The bill arrives in Congress at a time when US lawmakers remain divided over how aggressively the federal government should regulate AI.

The Trump administration argues that strict restrictions could undermine US competitiveness in AI, while some lawmakers and technology-policy groups are calling for tighter oversight of frontier models.

Reuters reported on the 22nd that US Secretary of State Marco Rubio has instructed US diplomats to rebut claims that American tech products contain "kill switches," while directing them to promote US AI products to foreign partners.

[News Distiller: distilling global information into what you really need to know.]