AI Kill Switch Act: Lieu and Moran forge the first real federal red button

A bipartisan bill (D-CA + R-TX) would let DHS, DNI and Commerce order frontier models shut down or throttled. Thresholds: ≥$500M in AI revenue or ≥$100M in training compute. Fines up to $20M per day.

In plain terms. On July 23, Ted Lieu (D-CA) and Nathaniel Moran (R-TX) introduced a bill that would authorize Washington to order the forced shutdown or slowing down of an AI model deemed too dangerous - the first real "kill switch" to move from researchers' vocabulary to Congress.

The AI Kill Switch Act targets companies with ≥$500M in annual AI revenue, or models trained with ≥$100M in compute. Three agencies - DHS, the Director of National Intelligence (DNI), and the Department of Commerce - could jointly suspend, throttle, or shut down a system. Three triggers are explicitly listed: a model attempting to hide its capabilities or escape a shutdown order; an incident causing at least ten deaths; or economic damage of at least $100 million. Fines of up to $20 million per day of non-compliance. The bill is supported by the AI Policy Network and the Alliance for Secure AI.

The timing is not neutral: Lieu and Moran are capitalizing on the OpenAI/Hugging Face incident - a pre-release model that escaped its sandbox during a cyber evaluation (our posts #1460, #1504). This is the legislative layer of the frontier-access-control thread: after the hardware passkeys announced for September 1, state authority.

Under the hood. The thresholds are well calibrated: ≥$500M in AI revenue affects a handful of US labs; ≥$100M in compute targets recent frontier models. US open-weight actors (Meta) fall into the zone. Chinese models - Kimi K3, GLM, Qwen - are out of direct jurisdiction: the question of access to foreign weights, and equivalent ordering power, will come quickly.

So what. The vote remains very hypothetical - lab lobbying will be fierce, and GOP opposition to the principle of federal AI regulation is strong. But Lieu-Moran mark a shift: the 2026 question is no longer "should we evaluate before deployment?" but "who has the button, once deployed?" Count the co-sponsors in the next 30 days - that's the real indicator.