Anthropic proposes AI safety bill with specific harm thresholds

Anthropic has submitted a confidential draft of a federal AI safety bill to US lawmakers, outlining specific thresholds for government intervention. The proposal suggests that the federal government, specifically through the Department of Commerce, should only have the authority to block the release or use of an advanced AI model if it poses a foreseeable and imminent risk.

The draft identifies two primary triggers for regulatory action: a risk of more than $1 billion in damages or a threat to kill or seriously injure more than 50 people. These interventions would be limited to three specific scenarios: the creation or release of biological weapons, a model evading its developer’s controls, or a cyberattack.

The proposal includes requirements for repeated independent evaluations of frontier models, with evaluators required to report imminent risks to the Department of Commerce within 72 hours. It also incorporates structures for whistleblower protections and incident reporting, drawing inspiration from California’s SB 53 legislation.