Anthropic's draft federal AI safety bill leaks and draws criticism from safety advocates

A discussion draft of a federal AI safety bill that Anthropic had been circulating to lawmakers became public on 16 October, when Politico published it, according to Transformer. The draft has no sponsor or bill number and has not been introduced. Several AI-safety advocates criticised it within hours.

The draft turns Anthropic’s June Advanced AI Framework into legislative language. As described by critics who read it, it requires third-party evaluators to assess frontier developers’ models and practices for catastrophic risk, with no licensing or standards for the evaluators. It would let the Commerce Department petition a federal court to block a dangerous model, with the court deciding “without deference” and weighing harm to the developer, users and the public. It would preempt state frontier-AI laws including California’s SB 53, New York’s RAISE Act and Illinois’s SB 315, with carve-outs for child safety, and it would bar state courts and legislatures from suspending a model on catastrophic-risk grounds. Transformer also credited it with improvements over the FRONTIER Act, including provisions on company security and on containment failure and automated AI research.

The criticism was that the draft was weaker than that bipartisan bill and than Anthropic’s own June framework. Encode’s Nathan Calvin said it would likely make a blocking order harder to obtain than under existing law. Shakeel Hashim of Transformer wrote that it “reads like Anthropic’s business interests winning out over its purported safety concerns”. Zvi Mowshowitz, who writes the newsletter Don’t Worry About the Vase, called it “unacceptable and troubling”. LawAI’s Charlie Bullock gave a more mixed verdict: he called it a “pretty reasonable AI safety bill” that was weaker than FRONTIER mainly on rulemaking and emergency shutdown authority, and better on auditor access.

On 18 October Anthony Cimino, Anthropic’s head of federal affairs, posted a thread calling the draft “a translation of those ideas into legislative language”, shared with “a small number of congressional offices”. He said Anthropic would revise its preemption language, defended requiring a meaningful standard before a court blocks a model, and said it was “still probably time to re-evaluate thresholds”. He added that Anthropic would publish proposals on independent evaluators and on preventing “evaluator shopping”. Mowshowitz replied that the response did not address what was centrally wrong with the draft.

Anthropic had endorsed SB 53 in 2025, which made the draft’s preemption of it the sharpest point of dispute. No revised draft had been published as of 20 October.