Anthropic Quietly Pitched Congress on Its Own Draft AI Safety Bill

What happened

The draft would let the government block a model's release or use only in narrowly defined cases: the risk must be foreseeable and imminent, causing more than $1 billion in damages, or killing or seriously injuring more than 50 people. Politico says the triggering risks are limited to three categories: the creation or release of a biological weapon, a model evading its developer's controls, or a cyberattack.

The draft sets up third-party evaluation: independent evaluators would register with the Department of Commerce and periodically assess labs and their frontier models; labs could pay these evaluators, as long as payment is not tied to the evaluation's outcome. If evaluators determine a frontier model poses an "imminent enumerated risk," they must report it to Commerce within 72 hours of completing their assessment.

On enforcement, the Commerce Department could request a judge to halt a model after determining risk; but the draft explicitly bars any other US government department, agency, or officer from suspending a model's development, deployment, or internal use. A federal judge for the District of Columbia could ultimately order a model's suspension.

Key facts

1. Shutdown bar: Foreseeable, imminent risk: over $1 billion in damages, or 50+ killed or seriously injured
2. Three triggers: Biological-weapon creation/release, model evading developer controls, cyberattack
3. Third-party evaluation: Independent evaluators register with Commerce and assess periodically; labs may pay them, but not based on outcomes
4. 72-hour reporting: Evaluators must report "imminent enumerated risk" to Commerce within 72 hours
5. Who can halt: Only Commerce may ask a judge to suspend a model; all other departments, agencies, and officers are barred from suspending development, deployment, or internal use
6. Reporting and response: Politico reported on October 16, citing three people with knowledge of the effort; Anthropic and Commerce did not respond to Reuters' requests for comment