US tech giant Anthropic blasts China’s top AI labs for stealing its model

Anthropic says three major Chinese AI labs ran large-scale, covert operations to copy its Claude model. In a public statement, the US firm accused DeepSeek, Moonshot and MiniMax of mounting "industrial-scale" distillation campaigns that generated more than 16 million interactions with Claude via roughly 24,000 fraudulent accounts. The company claims the labs used proxy networks and so‑called "hydra" clusters of fake accounts to bypass regional blocks and terms of service, concentrating traffic on Claude’s most valuable capabilities: reasoning, tool use and coding.

Distillation—training one model on the outputs of another—is common when labs apply it to their own systems, but Anthropic argues that using another company’s models this way is exploitation, not standard practice. It warns that models distilled illicitly are unlikely to preserve safety features, raising national security concerns if integrated into military, surveillance or cyber operations. The accusations echo earlier claims by OpenAI that DeepSeek copied ChatGPT, and come as US firms tighten security around data centres, internal networks and staff access. Anthropic says it is deploying technical detection tools and sharing intelligence with other AI companies and regulators, and is using the case to argue for tougher, coordinated responses and stronger export controls on advanced chips.