ME News reports that on September 11 (UTC+8), Anthropic’s latest report revealed that, over the past few months, it identified and blocked Claude distillation activities originating from seven Chinese AI companies: Alibaba, Moonshot AI, DeepSeek, Zhipu AI, Xiaomi, SenseTime, and MiniMax. While distillation itself is a common training method, Anthropic’s allegations focus on large-scale circumvention of restrictions, scraping of Claude outputs, and user data.
1. Alibaba: Largest in scale. Anthropic reported over 151 million interactions from May to July, peaking close to 3 million per day, primarily extracting the full reasoning processes of Opus 4.6 and 4.7 to train Qwen 3.5, 3.6, and 3.7.
2. The Dark Side of the Moon: Directly forwarding Kimi users’ queries to Claude. Nearly 300,000 real user requests were sent to Claude within 10 days, some including internal corporate code and valid credentials; related interactions exceeded 23 million between May and July.
3. DeepSeek: Also secretly forwards user requests to Claude, specifically identifying users who utilize tools such as Claude Code and OpenCode, then selectively forwards some requests to Opus. Over 12.1 million times in 14 days.
4. Zhipu: Specifically extracted and cleaned Claude’s reasoning process, and also trained its cybersecurity capabilities using U.S. models. Anthropic stated that Zhipu initially attempted to distill Fable but, after finding its defenses too strong, shifted to Opus 4.6 and a model from another U.S. company, achieving over 3.4 million interactions in 17 days.
5. Xiaomi: Feeds real conversations and programming logs generated by MiMo users back into Claude to generate SFT and reinforcement learning data. Over 400,000 requests in 20 days; Anthropic even suspected that the free trial period of MiMo-V2-Pro may have been used to attract more overseas developers to generate training data.
6. SenseTime: Purchases user chat logs with Claude directly from third-party data providers. These data originate from agents and intermediary services, and some users are unaware that their conversations are being stored and sold.
7. MiniMax: Accused of operating a model intermediary service through a company with no apparent connection, which only offers Anthropic and OpenAI models—excluding MiniMax’s own models. Anthropic believes this service may be designed to collect user interactions with U.S. models for training purposes.
All of the above are investigations, attributions, and allegations by Anthropic, none of which have been independently verified. (Source: Anthropic)