Anthropic reports nearly 200 million distillation exchanges tied to Chinese labs

A report Anthropic published on Thursday says it observed nearly 200 million exchanges across five separate distillation campaigns attributed to China-based companies. Distillation attacks aim to extract a model's chain of thought and then train a smaller model on those traces. Anthropic does not normally show that chain of thought to users; according to the report, attackers found prompts that coaxed the model into revealing the traces directly. The largest share was attributed to Alibaba: 151 million exchanges between May and July spread across 3,500 accounts, peaking at nearly three million a day, all sharing one fixed prompt. In a campaign linked to Moonshot AI, maker of Kimi, requests were said to be routed from the Chinese military — about 300,000 requests over ten days through 5,000 accounts, mostly aimed at the Opus model.