Anthropic proposes AI safety bill draft: independent pre-release evaluations, injunctions for 50+ deaths or $1B in property damage

The draft sets two trigger thresholds: more than 50 people at risk of death, or more than $1 billion in property damage, with the risk originating from bioweapon development, a model evading developer control, or cyberattacks. Evaluations would be done by independent evaluators registered with Commerce; AI labs may pay for them, but fees cannot be tied to results. Once an evaluator finds imminent catastrophe risk, it must notify Commerce within 72 hours, and Commerce then asks a D.C. judge to block release; the draft explicitly bars federal agencies from halting model distribution on their own. Like California's enacted SB 53, the draft reuses the 50-person and $1 billion thresholds. An Anthropic spokesperson confirmed to Politico that the draft stems from its Advanced AI Framework, lobbied to federal and state lawmakers since June. For AI labs planning U.S. frontier model releases, the draft means engaging Commerce-registered evaluators before launch and reserving a 72-hour reporting and court-application window; legal teams reviewing release compliance must reassess conflict-of-interest disclosure duties from paid third-party evaluations.