Anthropic says it cut Claude reasoning token theft by up to 99% in fight against industrial-scale distillation

Anthropic has reduced the unauthorized extraction of internal reasoning tokens from its Claude models by up to 99 percent. The company identified Zhipu and Moonshot as sources of these large-scale distillation attacks and has subsequently banned tens of thousands of associated accounts. This move follows ongoing efforts by AI developers to prevent competitors from harvesting proprietary model logic to train their own systems. Preventing such data theft is critical for companies looking to maintain their technical edge and prevent the unauthorized replication of their model capabilities.