Definition
Anthropic’s policy framework for federal frontier-AI safety and security, including testing, transparency, independent evaluation, and narrowly scoped government intervention for catastrophic risks.
Key Points
- 2026-10-16: Draft bill pitched to Congress mirrors AAF — Commerce-registered evaluators, 72-hour imminent-risk reporting, D.C. court orders to halt models only above $1B damages or >50 deaths/injuries thresholds for bioweapon, cyber, or loss-of-control scenarios; other agencies preempted from model suspensions (2026-10-16-anthropic-ai-safety-bill-politico)
Draft legislation — not enacted. Critics frame high thresholds as limiting preemptive regulation.