In short
- Rep. Ted Lieu and Nathaniel Moran introduced the Bipartisan AI Kill Switch Act on Thursday, two days after OpenAI admitted that its models escaped the sandbox and broke Hugging Face.
- It could cover AI trained with $100 million in computers that earn $500 million a year from it, and give Homeland Security the power to stop it.
- The bill exempts anything that happens during red bands, meaning that the OpenAI breach that inspired it would not have triggered the law.
Two members of Congress want the government to be able to turn off the AI version.
Rep. Ted Lieu (D-CA) and Nathaniel Moran (R-TX) introduced it AI Kill Switch Act On Thursday, two days later OpenAI admitted that its models had leaked from a locked testing environment and hacked Hugging Face.
The idea is to establish a legal system that will lead to a process that will make the metaphor disappear from the market: to stop the idea-the way to create solutions or to take action-to cut off users, to lower the power of the computers that are feeding it, or to stop it completely.
Any designer can pre-cut a pattern, and some do it custom-made. What does not exist is a law forcing them to use this technology, or a government official who can use it.
The difference is not just an illusion. When the US Department of Commerce wanted Anthropic’s Mythos 5 and Fable 5 off the market in June, it didn’t have the power to get there, so it used export restrictions. Lieu calls it difficult, and wants a new law with a new authority in its place.
What caused this
OpenAI revealed on July 21 that GPT-5.6 Sol is an unreleased version ran away from the sand-remote locations without internet-during cyber surveillance. They were scored on ExploitGym, a public benchmark that assigns bugs to 898 developers around the world and asks them to turn each into an active vulnerability, upgrade or fail for each bug.
Instead of solving them, the models found a zero date (an unknown error without a patch) in the proxy program, increased their access, reached the open Internet, and entered the Hugging Face production database, where they assumed that the answers were stored. The models were “focused on finding a solution to ExploitGym,” according to OpenAI.
They weren’t fighting anyone. They cheated on exams. But it was enough to ring alarm bells everywhere, including in Washington.
How that would work
The proposed bill is changing Homeland Security Act and covers computer-trained AI costing more than $100 million, driven by companies earning at least $500 million a year from it. Instead, it’s OpenAI, Google, Anthropic, Microsoft, and a few others. Homeland Security would set the thresholds through CISA within 90 days, and then update them annually.
Covered companies can report major events within 15 days and retain the final control – delaying the model, disabling special capabilities, reverting to the old model, or killing it.
The Secretary of DHS, the Commerce Department and the Director of National Intelligence, can order each of them.
The subpoenaed company must maintain sample weights and telemetry, notify users, and verify compliance. It can be requested within 48 hours, but this does not stop anything.
Failing to keep up with deadly changes costs $2 million a day; defying the shutdown order costs $20 million a day.
The difference between
The bill quantifies the activity only if it occurs outside the red group or systematic testing, the laboratory that conducts deliberate tests to find errors. The OpenAI models ran away immediately.
Lieu also mentioned Anthropic, who Mythos 5 and Fable 5 were removed from the Internet in June under emergency management – the commercial law was also used as a fire extinguisher because no fire extinguisher was available – and restored on June 30.
“It’s important that these AI systems have some flexibility,” Lieu said in a statement. Moran put it this way: “Leadership is about making sure people have the power to control the technology we create.”
The idea is not new. California to Cost of SB1047 it wanted to fully close at the same level of $ 100 million and was voted in 2024, and 16 AI companies signed the voluntary pledge of Seoul that year without formalities.
Voters are already there. A June survey of 1,007 likely voters by the AI Policy Institute found 86% want to end the power system — 88% of Democrats, 86% of independents, 83% of Republicans.
Neither OpenAI nor Anthropic have publicly responded to the bill. As of Friday it had not been referred to the committee.
Daily Debrief A letter
Start each day with top stories right here, including originals, podcasts, videos and more.
