A leaked bill that Anthropic has been urging on lawmakers would let the federal government stop an AI model only after a narrow kind of risk, at a high bar. Tom's Hardware says the trigger is 50 deaths or $1 billion in damages, and that the draft would leave the Commerce Department as the agency that can take a model to a judge.

Tom's Hardware's account and PC Gamer's both quote an Anthropic spokesperson. "This bill represents the proposals outlined in Anthropic's Advanced AI Framework, which we have been advocating for with federal and state lawmakers since June." The spokesperson called them sensible policies Congress should act on. PC Gamer noted the resemblance, drawn in the piece it was following, to California's AI statute.

The covered events are a bioweapon, a cyberattack, or a lab losing the model

The bill, as Tom's Hardware describes it, limits the risk to three categories: the creation or release of a bioweapon, an AI-enabled cyberattack, or a lab losing control of a model or its agents. If one of those is about to happen, the evaluator has 72 hours to report it to the department. The department then has to put the evidence in front of a federal judge in the District Court for the District of Columbia. Below that threshold, this draft does not hand the government a stop button.

Anthropic tested a worm that copies model weights onto cheap hardware. That test is one reason the company has talked about loss of control. Six METR researchers now have employee-level access inside Anthropic. The leaked bill is the statute those evaluators would be reporting under, if Congress passed it. METR had already raised about $71 million as the labs hired auditors. The bill is a different instrument from the funding.

Amodei's slowdown pitch is the politics around the text

Tom's Hardware ties the draft to a warning from chief executive Dario Amodei about a "swarm" of AI botnets in the next 6 to 12 months, and to a proposed slowdown that Sam Altman and Elon Musk have said they agree with. The same piece says Congress is already considering kill switches on the most powerful models, and that Senator Bernie Sanders has proposed a 20-year prison term for developing what he calls AI superintelligence. Those are other bills. Anthropic's text is the one with the 50-death and $1 billion gates, and with Commerce as the door to the court.

PC Gamer's commentary is that a threshold this high lets a company look like it is accepting rules while the rules only bind it on terms the company helped write. The piece it followed also set the draft next to California's AI statute. Tom's Hardware adds two other proposals already in the conversation: a congressional push for kill switches on the most capable models, and a Bernie Sanders bill that would impose a 20-year prison term for building what he calls AI superintelligence. Neither of those is the leaked text. The leaked text is the one with a death count, a dollar count, a 72-hour clock, and a single department in front of the federal court in Washington. It is not law.