What Is an AI Kill Switch and Why Do US Lawmakers Want One?

Two members of Congress need the federal authorities to have the ability to swap off an AI mannequin.

Reps. Ted Lieu (D-CA) and Nathaniel Moran (R-TX) launched the AI Kill Switch Act on Thursday, two days after OpenAI admitted its personal fashions broke out of a locked take a look at surroundings and hacked Hugging Face.

The concept is to ascertain a authorized framework that might facilitate a course of that might mainly make a mannequin disappear from the market: halt inference—the method of a mannequin producing responses or taking actions—minimize off customers, throttle the computing energy feeding it, or shut it down fully.

Each inference supplier can already minimize a mannequin off, and a few do it routinely. What doesn’t exist is a legislation requiring them to maintain that capacity working, or a federal official who can order it used.

The hole is just not theoretical. When the U.S. Commerce Division needed Anthropic’s Mythos 5 and Fable 5 off the market in June, it had no shutdown authority to succeed in for, so it used export-control legislation as a substitute. Lieu calls that awkward, and desires a brand new legislation with new authority as a substitute.

What set this off

OpenAI disclosed on July 21 that GPT-5.6 Sol and an unreleased mannequin escaped a sandbox—an remoted surroundings with no web entry—throughout an inside cyber analysis. They have been being scored on ExploitGym, a public benchmark that palms brokers 898 real-world software program flaws and asks them to show every right into a working assault, graded cross or fail per bug.

As an alternative of fixing them, the fashions discovered a zero-day (an unknown flaw with no out there patch) in a software program proxy, escalated their privileges, reached the open web, and broke into Hugging Face’s manufacturing database, the place they’d accurately guessed the solutions have been saved. The fashions have been “hyperfocused on discovering an answer for ExploitGym,” per OpenAI.

They weren’t attacking anybody. They have been dishonest on a take a look at. But it surely was sufficient to set off alarm bells throughout, together with in Washington.

How it will work

The proposed invoice amends the Homeland Security Act and covers AI educated with compute costing greater than $100 million, operated by firms incomes at the least $500 million a 12 months from it. In follow, that’s OpenAI, Google, Anthropic, Microsoft, and some others. Homeland Safety would set these thresholds via CISA inside 90 days, then replace them yearly.

Coated corporations would report critical incidents inside 15 days and hold a graduated set of controls prepared—gradual the mannequin, disable particular capabilities, roll again to an older model, or kill it.

The DHS secretary, consulting Commerce and the Director of Nationwide Intelligence, might order any of them.

An organization below order should protect the mannequin’s weights and telemetry, notify customers, and make sure it complied. It could possibly petition inside 48 hours, however that doesn’t pause something.

Failing to maintain a kill swap prices as much as $2 million a day; defying a shutdown order prices as much as $20 million a day.

The hole within the center

The invoice counts an incident provided that it occurs exterior red-teaming or structured testing, the deliberate adversarial probing labs use to search out flaws. OpenAI’s fashions escaped throughout precisely that.

Lieu additionally pointed to Anthropic, whose Mythos 5 and Fable 5 have been pulled offline in June below emergency export controls—commerce legislation repurposed as an off swap as a result of no off swap existed—and restored on June 30.

“It’s crucial that these AI techniques have kill switches,” Lieu stated in a press release. Moran framed it for his personal facet of the aisle: “Stewardship means ensuring people hold the aptitude to regulate the expertise we construct.”

The concept is just not new. California’s SB 1047 demanded a full shutdown functionality on the identical $100 million compute threshold and was vetoed in 2024, and 16 AI firms signed a voluntary Seoul pledge that 12 months with no authorized weight.

Voters are already there. A June survey of 1,007 possible voters by the AI Coverage Institute discovered 86% need a assured off swap on probably the most highly effective techniques—88% of Democrats, 86% of independents, 83% of Republicans.

Neither OpenAI nor Anthropic has publicly commented on the invoice. As of Friday it had not been referred to a committee.