Anthropic says defense measures cut stolen Claude reasoning tokens by 90% to 99%

Anthropic states its new defense measures reduced the number of extracted reasoning tokens from Claude by 90% to 99%, guarding against industrial-scale distillation attacks targeting its models. Attackers used hundreds of thousands of fake accounts to extract the model's internal chain-of-thought steps; Anthropic implicated Zhipu and Moonshot, stating Zhipu stole the web capabilities of leading US models for GLM-5.3, while Moonshot secretly forwarded queries through Claude.