Anthropic details distillation campaigns from Alibaba, Moonshot AI, and DeepSeek — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic released a report Thursday alleging persistent distillation attacks by China-based AI companies, identifying five separate campaigns totaling nearly 200 million exchanges aimed at extracting Claude's chain-of-thought reasoning for use in training smaller models.
- Alibaba ran the largest campaign Anthropic has ever observed — 151 million exchanges between May and July 2026 across 3,500 accounts, peaking at roughly 3 million exchanges per day, all sharing a single fixed prompt used to harvest training material for Alibaba's Qwen model family.
- Moonshot AI, maker of Kimi, operated a separate campaign that Anthropic says routed requests directly from the Chinese military, including one query asking Claude to assess closed-circuit surveillance footage for "abnormal behavior" — nearly 300,000 requests over 10 days via 5,000 accounts primarily targeting Anthropic's Opus model.
- Attackers bypassed Anthropic's standard "summarized thinking" blocks by framing queries as translation requests — e.g. instructing Claude to translate its working memory into "katakana-only Japanese" — to trick the model into revealing raw thinking traces.
- OpenAI has previously reported similar activity it attributed to DeepSeek, and Anthropic itself flagged distillation attacks back in February, but the company says the new campaigns are both larger and more aggressive than prior efforts.
Why it matters: Alibaba's 151 million exchanges represent the largest wholesale distillation effort Anthropic has documented, suggesting Chinese labs are harvesting frontier-model reasoning at industrial scale rather than building from scratch. The Moonshot campaign's routing of Chinese military surveillance queries through consumer accounts blurs the line between commercial AI competition and state intelligence collection.
Ask SkimNews
