Anthropic releases detailed AI safety framework

The expanded plan builds on recent comments from Dario Amodei about authorities’ right to block dangerous models. The proposals focus on systems trained with over 10^25 floating-point operations, where the company invests over $1 billion annually or earns $500 million+.

Key requirements include mandatory risk reporting and third-party assessments. Anthropic advocates for defenses against cyberattacks on model weights and training infrastructure. The company also proposes tracking model distillation attempts.

Beyond developer controls, the document offers recommendations on societal resilience, including screening gene synthesis to prevent biological weapon creation.