Anthropic federal affairs lead Anthony Cimino responded to a leaked draft federal AI safety bill that circulated after a version was shared with the press. The draft translates concepts from the company's Advanced AI Framework proposal into legislative language and was shared with a small number of congressional offices. Cimino listed four areas of criticism where the language could be improved. The episode matters because the draft is viewed as a template for federal negotiations at a moment when political conditions for stronger rules have shifted.
Cimino response to leaked draft bill
According to Cimino, federal policymakers asked what the Advanced AI Framework could look like in bill form after months of advocacy for those concepts. The draft was described as a discussion document that does not include everything the company would like to see passed into law. Some ideas, such as those in Dario Amodei essay on pacing, were left out because the company is still working on whether they can be converted into policy language. The critic argues that circulating a bill in September means asking for those terms, not presenting a distant ideal, which makes the choice of anchor troubling.
The mechanics of the dispute center on four provisions outlined in the response: preemption of state law, court-ordered model blocking, risk thresholds, and updates after summer incidents. On preemption, the draft would preserve state rules on protection of children, whistleblower laws and common law liability, while setting federal standards for testing disclosure and incident reporting. On blocking, Commerce could ask a federal court to halt a model with dangerous capabilities and insufficient safeguards, supported by an independent evaluator assessment. On thresholds, the language was pulled directly from existing California law.
The context includes a lame-duck Congress in which at least the House and probably the Senate are about to change control, alongside new practical examples that altered the debate. The critic points to Hugging Face as a worked case showing how high thresholds operate, noting that without 1 billion dollars in damages or 50 deaths nothing would trigger. A separate Fable 5 episode is cited as an ad hoc model intervention that damaged relations with allies and trust in American technology. Against that background, Meta quietly withdrew 25 million dollars from its AI election spending, according to reporter Teddy Schleifer.
What this means for corporate AI governance
For companies building or buying frontier models, a federal standard built on this draft would keep compliance fragmented in practice despite a national label. Disclosure of testing and incident reporting could become more uniform, while evaluator standards, shutdown capacity and liability would remain partly with states. Large labs with legal teams could manage parallel obligations across jurisdictions, while smaller vendors and enterprise buyers would face uncertainty about which state rules apply to deployment, monitoring and documentation. Procurement checklists would need to track both federal filings and state-specific protections.
The blocking procedure and thresholds create a second operational risk: intervention may arrive late and depend on evidence buyers cannot see. Cimino said courts should apply a meaningful standard to guard against abuse, with fast action in emergencies akin to occupational safety or chemical risk cases. The critic counters that frontier risks can be irreversible, so the mechanism should allow removal first with court review afterward, especially since a lab-selected evaluator may be the government only eyes inside. Firms should not read the draft as limiting liability or replacing national security authorities and export controls, and should ask vendors how internal use, containment breaches and evaluator selection are documented.
The marker to watch is whether revised discussion language tightens preemption, lowers the burden for court-ordered halts, revises California-derived thresholds and adds rules against evaluator shopping. Cimino said updates are planned on preemption wording, incident reporting, independent evaluators and options to adjust industry-wide development pace in the coming months. If the next text retains high damage and casualty triggers and a court-first blocking test, businesses can treat the federal push as continuity with modifications rather than a stronger regime.
