Anthropic Accuses Alibaba, DeepSeek of Massive Claude Data Theft via 25K Fake Accounts

Anthropic has formally alleged that Alibaba's Qwen lab orchestrated an industrial-scale distillation attack against its Claude model, using over 25,000 fake accounts to extract 28.8 million question-and-answer pairs between April 22 and June 5. The company presented this to Capitol Hill as the largest known instance of model distillation — a technique where outputs from a frontier model are collected to train cheaper alternatives. Earlier in February, Anthropic had already accused DeepSeek, Moonshot AI, and MiniMax of similar patterns. The White House Office of Science and Technology Policy in April publicly characterized such activity as industrial-scale theft of American AI intellectual property, indicating growing policy concern.

OpenRouter data cited by Forbes shows that Chinese-developed models now represent roughly 61% of token consumption among the top 10 on the platform, up from under 1.2% in late 2024 — meaning Western developers on a Western routing layer are actively choosing these distilled alternatives. The allegations underscore a shift in the US-China AI competition narrative: while chip export controls aim to limit compute access, the more consequential story may be about training data. However, claims remain unproven in court, and distillation's exact contribution versus open-weight practices or original Chinese research remains unclear. The forward-looking tension is whether US labs tighten API terms and detection or concede the middle market, while downstream startups benefit from cheaper, near-frontier models regardless.