Anthropic plans changes to draft AI safety bill after criticism

Anthropic plans to revise provisions in its draft federal AI safety legislation after safety advocates criticized restrictions on stopping dangerous models and overseeing independent evaluators. Anthony Cimino, who leads the company’s federal affairs team, said on October 18, 2026, that language concerning state laws needed clarification.

The company has circulated the proposal among congressional offices. According to Politico, which published the confidential draft, Commerce would need a federal court’s approval to block a model. Qualifying risks would have to be foreseeable and imminent, involving damage exceeding US$1 billion or more than 50 people being killed or seriously injured. They would also need to stem from a bioweapon, a cyberattack or a model escaping developer control.

The proposal would prevent other agencies or officials from suspending models and override some state requirements concerning disclosure, outside evaluations and incident reports. Evaluators would register with Commerce and notify it of imminent risks within 72 hours. Labs could pay evaluators, provided compensation was not tied to their findings.

Writing for Transformer, Shakeel Hashim argued that the draft offered weaker protections than the bipartisan FRONTIER Act, introduced in July by Representatives Jay Obernolte and Lori Trahan. He cited the absence of evaluator licensing and standards, alongside greater obstacles to blocking models. California State Senator Scott Wiener also criticized its proposed safety requirements.

Cimino said Anthropic would preserve state roles on evaluator standards and labs’ ability to shut down models. He said the company planned additional evaluator proposals and was considering ways to prevent labs from shopping for favorable assessments. He also said risk thresholds probably needed reconsideration.