Anthropic drafts federal AI bill letting a court block frontier models that risk 50+ deaths or $1B in damage

This is a policy proposal, not a model release, so no model or version is involved. Anthropic's framework, as summarized by a third party, would cover developers that train models above 10^25 FLOP and also either earn more than $500M a year in AI revenue or spend more than $1B a year on AI R&D (regulations.ai). Public reporting does not say whether the draft bill uses the same thresholds. The bill text is not public.

• Evaluation design: Evaluators would register with Commerce and periodically assess developers and their models. Labs could pay them, as long as payment does not depend on the results (Interesting Engineering).
• Framework depth: The framework calls for at least one independent review of risk assessments within six months, with access to unredacted reports and model weights. It also calls for six-monthly risk reports and reporting of critical safety incidents within 15 days (regulations.ai).
• Gaps: Public reporting describes no test protocols, capability benchmarks or method for judging "imminence." How evaluations would be run or reproduced is undefined.

• The thresholds and the three risk categories rest on Politico's reading of a confidential draft. Other outlets repeat it rather than confirm it independently (Reuters via Investing.com).
• Anthropic confirmed the draft and said it reflects the framework it has promoted since June (TechCrunch).
• Reports describe the 72-hour rule differently. TechCrunch says evaluators get 72 hours to notify Commerce after finding an imminent risk. Interesting Engineering says only that results must be reported within 72 hours (Interesting Engineering).

• SB 53: The bill borrows SB 53's 50-person and $1B thresholds. SB 53 itself relies on disclosure and incident reports to California's Office of Emergency Services; it has no court-ordered halt (FPF).
• Competing bills: These include the FRONTIER Act, a Thune/Klobuchar/Cruz Senate effort, and Casar/Sanders proposals for development pauses (Interesting Engineering). Senator Warner's S. 5576 sets broader safety requirements without comparable intervention thresholds (Law Commentary).
• Other industry positions: Nvidia's Jensen Huang has argued that no new AI laws are needed (TechCrunch).

• Lab-paid evaluators: Both TechCrunch and Law Commentary flag a conflict of interest when labs pay their own evaluators (Law Commentary).
• Limits on government power: The draft bars any agency from halting a model on its own, which critics read as narrowing regulators' authority (Interesting Engineering).
• State preemption: The draft reportedly overrides some state requirements on risk disclosure, evaluation and incident reporting. Anthropic's own framework says Congress should preempt state law only with a federal regime at least as strong (regulations.ai). Whether the draft meets that bar has not been assessed publicly.
• Scope: TechCrunch notes that agentic-hacking incidents are a more immediate industry concern than mass-casualty scenarios.

• One federal standard: Frontier labs would face one federal regime instead of several state laws, but would have to fund third-party evaluations and run formal security programs.
• High bar for a halt: Model releases could be stopped only through a court case meeting high thresholds, which makes sudden regulatory stops less likely.
• New evaluator market: A market for Commerce-registered evaluators would emerge. Their credibility would depend on rules for independence that have not been published.
• Status: Treat this as one of several competing proposals, not pending law.