Anthropic has circulated a draft AI safety bill to members of Congress that would let the government block a model's release only if it risks causing more than 50 deaths or injuries or $1 billion in damage, Politico reported. The draft names the Commerce Department as the only federal body able to order a lab to halt a release, acting through a federal judge in the District Court for the District of Columbia. Independent evaluators registered with Commerce would periodically assess labs and their models and report any imminent risk within 72 hours.
The covered risks are narrow and specific: the creation or release of a bioweapon, AI-powered or AI-enabled cyberattacks, or a lab losing control of a model or its agents. Once an evaluator flags one of those scenarios, the department would have to put its evidence before the DC court, which would decide whether to stop the release. The draft bars every other department, agency or officer from issuing such an order. It also allows labs to pay the evaluators assessing them, provided the payment does not compromise their independence.
Anthropic confirmed the document reflects its own agenda. "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," a spokesperson told Politico, calling them "sensible policies we believe Congress should act on." Bloomberg and Politico both characterised the circulation as deliberately quiet, a contrast with louder industry coalitions.
That quietness is itself a tactic. Sharing finished statutory text, with sections, definitions and enforcement hooks, lets a company set the terms of debate: which risks appear first, how a "frontier model" or "deployment" is defined, and whether open-weight releases are treated differently. Those definitions ripple into compliance costs, insurance underwriting and product roadmaps. Venture-backed startups tend to fear that fixed compliance costs entrench incumbents, while incumbents argue a single federal rule beats a patchwork of state laws.
The draft lands amid a run of safety alarms and competing proposals. Anthropic chief executive Dario Amodei has warned of a "swarm" of AI-driven botnets emerging within 6 to 12 months, and the company has cited existential risk as a factor in its IPO prospectus. AI Daily reported on 17 October that Anthropic had begun embedding third-party evaluators inside its own operations, a private version of the oversight this bill would codify. In Congress, a separate proposal would mandate kill switches on the most powerful models, and Senator Bernie Sanders has floated a 20-year prison term for anyone building superintelligence.
Against that backdrop, the thresholds are the contested part. Critics quoted by Tom's Hardware argued that a bar set at 50 casualties or $1 billion would leave most harms unaddressed and that a single US agency cannot reach a lab operating from another country. Supporters counter that a clear federal trigger beats the current vacuum. What the draft does not resolve is whether Congress will adopt Anthropic's thresholds, rewrite them in committee, or pass one of the rival bills instead. For now it is a starting position, not a law.