Anthropic is quietly circulating a draft federal framework on Capitol Hill that would let the government halt a frontier AI model only when it threatens more than $1 billion in damages or more than 50 deaths or serious injuries, according to reporting from Reuters, POLITICO and TechCrunch. Under the draft, the U.S. Department of Commerce would be the only federal body allowed to seek a court order against a model's development, deployment or release, and it could not act alone: it would have to petition a judge in the U.S. District Court for the District of Columbia, who alone could grant the suspension. Every other federal agency would be explicitly barred from independently stopping a model, even one it believed dangerous.
The thresholds are narrow in a second way. The qualifying harm must fall into one of three categories: a cyberattack, a model evading its developer's own controls, or the creation or release of a biological weapon. A model that enabled large-scale fraud below the billion-dollar line, or caused dozens of injuries short of 50 deaths, would not trigger the mechanism as reported. The draft pairs that blocking power with disclosure duties running the other way, requiring companies to test their own models and publish the results, and setting up a network of independent evaluators, registered with Commerce, who would have 72 hours to notify the department after spotting a qualifying risk.
None of this is law, or even a bill. No lawmaker has introduced the text, no committee has scheduled a hearing, and no vote exists. It is a lobbying document, and the gap between what it proposes and how it is being received is the story. Mother Jones treated the numerical bar itself as the headline, arguing that a threshold measured in deaths sets an unusually high floor for action. POLITICO focused on the mechanics of why the thresholds sit where they do. Critics zeroed in on a structural point: because Commerce would have to assemble evidence that a risk is both foreseeable and imminent before approaching a judge, the design invites a wait-and-see posture, with a single agency and a single court as the entire federal safety valve.
The draft reflects Anthropic's commercial position as much as its safety brand. The company has spent roughly two years arguing that frontier AI carries catastrophic risk, and its own Advanced AI Framework already sets internal thresholds for delaying a model's release. A single, predictable federal standard is easier for any lab to plan product launches around than the patchwork now forming: California and Colorado have advanced state-level rules, and the European Union continues phasing in its broader, risk-tiered AI Act. The timing also sits awkwardly against Anthropic's push to ship autonomous agents built on Claude Opus 5.5, a company racing to deploy more capable systems has an obvious interest in knowing exactly where the regulatory tripwires sit. Reporting has not documented an on-record response to the specific draft from OpenAI, Google or Meta.
Public skepticism frames how the numbers land. A 2025 Pew Research Center survey found 62 percent of U.S. adults had little or no confidence in the federal government to regulate AI effectively, which helps explain why a bar written by an AI company reads as generous to one audience and alarmingly narrow to another. What happens next is unsettled. Most industry-pitched frameworks circulate for a cycle and die without a sponsor, but the volume of coverage the $1 billion and 50-death figures have already drawn raises the odds that some version of the thresholds, the single-court design or the evaluator network resurfaces in a bill someone eventually introduces.