We define “AI Safety Levels”, a way to classify an AI system’s potential for catastrophic misuse and autonomy, and commit to safety and security measures commensurate with risk, along with evaluations for new training runs and deployments.

→ View original post on X — @anthropicai,