Anthropic pitches a U.S. AI safety bill that could block risky models

Lives at risk, as a threshold

50+ people, or $1 billion in damage

A model's release could be blocked over a foreseeable, imminent risk to more than 50 lives or more than a billion dollars in property damage.

Covered risks

• A biological weapon
• A model evading its developer's controls
• A cyberattack

The risk would have to come from one of three sources.

Independent evaluators would have 72 hours to alert the Commerce Department, which would ask a judge to halt the release. No federal agency could halt a model on its own.