Alibaba + DeepSeek : Almost 200 Million queries
Anthropic has revealed 5 instances of Claude Distillation by Chinese firms, which attempted to obtain ~200M conversation exchanges, with Alibaba attempting up to 3M Qwen training queries/day.
Anthropic dropped an intel report on the 5 colluding instances of Claude distillation by 3 Chinese firms, which stole almost 200 million conversation exchanges. Alibaba went after 151M exchanges over 3,500 accounts between May–July 2026, at up to 3M queries/day to train their Qwen models. Moonshot AI went after 5,000 accounts, and some of the queries were reportedly traced back to Chinese military sources (Anthropic)
Claude distillation is the equivalent of chain-of-thought exfiltration. The bad actors steal the reasoning process of a model, and replicate it in a smaller model of their own. With 5 instances in 6 months, it seems that this has turned from a few individuals to a full-blown operation. How can closed source models possibly stay secure?
Source: techcrunch.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →▌ Related guides
Author
Evgenii Arsentev
PhD · Chief Executive Officer, digital health
Articles · Latest articles