Strengthening our defenses against distillation attacks

Anthropic has significantly reduced the effectiveness of distillation attacks on Claude, decreasing harvested data by 90-99%. For instance, Zhipu's GLM-5.3, despite distilling US models, switched from Claude Fable 5 after encountering strong cyber safeguards. Anthropic has implemented stricter usage limits and account bans for persistent distillation attempts, and since September 24, charges for blocked requests. These new defenses, including a two-stage classifier, are believed to reliably prevent many more attempts to distill their frontier models.