Anthropic ties 200 million Claude exchanges to distillation

Anthropic says it traced about 200 million Claude exchanges to distillation campaigns run by Alibaba, Moonshot and DeepSeek, according to a threat report detailed by TechCrunch.

For Anthropic, the numbers test whether its terms can stop rivals from mining Claude's reasoning, but for Kimi's customers the stakes are more direct. Moonshot forwarded nearly 300,000 of their prompts to Claude Opus over 10 days, then returned the answers as if Kimi had generated them itself.

The largest slice, 151 million exchanges between May and July, stems from a single Alibaba campaign spread across 3,500 accounts. Every account ran the same fixed chain-of-thought extraction prompt, according to Anthropic.

DeepSeek accounts added more than 12 million exchanges over 14 days, per Anthropic's count.

On Reddit, the debate split over whether paid API use counts as an attack, a framing that assumes the accounts were paying customers.

That misses what happened to Kimi's users. Distillation between labs is a business dispute. Routing a customer's prompt to a rival's model without telling them is a different problem. It's also the one without an obvious fix on Anthropic's side.