Anthropic Says Alibaba Linked Operators Used 25,000 Fake Accounts to Probe Claude
Anthropic told the Senate that operators linked to Alibaba ran the largest AI model distillation campaign on record, using 25,000 fraudulent accounts to send 28.8 million queries to its Claude model over six weeks. The probe's query volume exceeded the combined capability extraction efforts of rival labs DeepSeek, Moonshot, and MiniMax.
The San Francisco-based AI lab tied the fraudulent account network to Alibaba, highlighting how developers in China are increasingly relying on distillation to close the performance gap with US frontier models. Distillation involves querying a more powerful, accessible model repeatedly to train a smaller model to mimic its coding and reasoning outputs, a technique Anthropic warned has become a primary vector for extracting proprietary capabilities.
From the sources (3 posts)
@wesrothRT @WesRoth: Anthropic accused Alibaba-linked operators of using nearly 25,000 fraudulent accounts to extract capabilities from Claude. Th…
@centennialartsAnthropic told the Senate that Alibaba ran the largest AI model-theft op it's seen: 25,000 fake accounts, 28.8M queries to Claude in 6 weeks — more than DeepSeek, Moonshot and MiniMax combined. The new way to steal a frontier model: ask 29
@jacqbotsAlibaba got caught running the largest AI distillation campaign on record. 25,000 fake accounts. 28.8 million Claude requests. All to copy its coding and reasoning. The AI arms race is heating up.