Anthropic proposed its own US law on frontier AI safety

Anthropic confidentially shared a draft federal law with US lawmakers that would allow stopping the deployment of a frontier model only in the case of a predictable, inevitable catastrophic risk.

Compiled by Sergey KostenchukPublished 2026-10-16Updated 2026-10-17

2026-10-16 Safety OpenAI

🛡 Anthropic proposed its own US law on frontier AI safety

The developer of Claude quietly showed Congress a draft federal law, Politico reports. Deployment can be stopped only in the case of a predictable, inevitable catastrophic risk: damage of $1 billion or more or serious harm to the health of more than 50 people. Scope: bioweapons, loss of control, cyberattacks.

🌍 If the draft becomes law: mandatory independent testing of frontier models, registration of assessors with the Department of Commerce, and 72-hour risk reports; compliance for OpenAI and Google will become more expensive.

👤 This is an early draft of an "emergency brake": only a judge can stop a model at the request of the Department of Commerce. It has not yet been introduced in Congress.

Source 1: https://www.politico.com/news/2026-10-16/anthropic-ai-safety-01124654 Source 2: https://www.reuters.com/legal/litigation/anthropic-quietly-pitches-its-own-ai-safety-bill-lawmakers-politico-reports-2026-10-16/

Sources