Anthropic accuses Chinese rivals of industrial‑scale Claude data theft as IBM stock

Anthropic disclosed that three Chinese AI labs—DeepSeek, Moonshot and MiniMax—executed coordinated “distillation” attacks on its Claude model, according to a report by The‑Decoder. The labs created more than 24,000 fraudulent accounts and submitted over 16 million queries, targeting Claude’s reasoning, programming and tool‑use capabilities.

DeepSeek alone generated roughly 150,000 queries aimed at extracting Claude’s chain‑of‑thought reasoning, reward‑model data for reinforcement learning, and censorship‑compliant answers on politically sensitive topics. Moonshot contributed about 3.4 million exchanges focused on agent‑based reasoning, computer‑vision analysis and reconstructing Claude’s thought processes. MiniMax accounted for the remaining 13 million queries, concentrating on agent‑based programming and orchestration, and reportedly pivoted to a newer Claude version within 24 hours of detection, The‑Decoder added.

All three labs employed proxy services to bypass China’s internet restrictions, allowing them to access Claude’s API at scale. Anthropic said the volume and diversity of the queries indicate an industrial‑scale effort to train weaker, downstream models using Claude’s outputs, a practice it deems a direct violation of its terms of service.

The allegations come as Anthropic’s Claude recently announced a new capability to automate COBOL modernization, a move that sent IBM shares tumbling, reported by Zerohedge. Bloomberg noted the development could disrupt legacy‑system maintenance across finance, airlines and government, amplifying the stakes of the data‑theft claims.

Anthropic has not yet filed a formal complaint, but the company warned that the attacks could compromise its competitive edge and urged policymakers to consider stricter enforcement against cross‑border AI data harvesting, per The‑Decoder.