AI Kill Switch Act: Congress's shutdown proposal hits a technical problem
A bipartisan bill would give US authorities kill-switch power over frontier AI. But critics point to a sweeping weakness: a regulator cannot switch off open model weights that have already been downloaded locally — and Congress likely…
AI Kill Switch Act: Congress's shutdown proposal hits a technical problem
A bipartisan bill would give US authorities kill-switch power over frontier AI. But critics point to a sweeping weakness: a regulator cannot switch off open model weights that have already been downloaded locally — and Congress likely won't act before the election anyway.
The bipartisan proposal
According to NBC News (September 14, 2026), Representative Ted Lieu (D-CA), co-chair of the Democratic Commission on AI, has partnered with Representative Nathaniel Moran (R-TX) on what they call the AI Kill Switch Act.
What the bill would actually contain is unclear: no full legislative text is publicly available. According to a review by 24ai.no — which for now is the primary source for its contents — the proposal would give the Department of Homeland Security authority to order developers of frontier AI models to restrict, pause, or shut down systems deemed to pose a catastrophic security risk. This account cannot yet be confirmed against primary sources.
The criticism: no switch when the model sits in a basement
The most sweeping technical objection concerns not the bill's policy, but its reach. Rob T. Lee, research lead at the SANS Institute, cited in 24ai.no, points out that large open-weight models can be downloaded and run locally — including on machines with no network connection:
"This is not a hole in the fence. On the open side, there is no fence. No regulator can reach into a basement in another country and switch off a file."
The point is simple: a "kill switch" can work against systems a developer itself operates, such as centralized API services. But when model weights can be freely copied, there is no technical mechanism to pull them back. The law's effect then depends entirely on it reaching what runs within the regulator's reach — and only that.
History repeats itself: SB 1047
The objection is not new. In California, a similar pattern emerged in 2024 with SB 1047, which originally required shutdown mechanisms for large models. The proposal was amended along the way, but Governor Gavin Newsom vetoed it anyway in September 2024, citing that regulation based on model size alone was the wrong approach (review in 24ai.no).
Political gridlock
Even if the bill were to move forward, no action is likely forthcoming. NBC News writes that Congress, despite new warnings, is unlikely to do anything before Election Day, even as experts say the technology grows in capacity every month.
President Trump reinforced the political resistance on Monday, September 14, when he wrote in a series of social media posts: "AI taking over the World, destroying Humanity, and all other things bad, is a HOAX" — cited via Yahoo News.
What remains
The proposal is still before committee, and no public legislative text exists for either the AI Kill Switch Act or what 24ai.no refers to as the "Stop Rogue AI Act." Central questions remain open: How far should the authorities' reach extend, how should "catastrophic security risk" be defined — and how should the law deal with models that have already been downloaded in thousands of places? Until a text is available, the answers are unknown.