Anthropic circled the Chinese big model: 7 companies distilled Claude's technique exposed According to Anthropic's latest report, Claude distillation activities from seven Chinese AI companies have been discovered and banned in the past few months. Ali, the dark side of the moon DeepSeek、 Zhipu, Xiaomi, SenseTime, and MiniMax were all named. Distillation itself is a common training method, and Anthropic's focus is on large-scale bypassing of restrictions, capturing Claude's output, and user data. 1. Alibaba: The largest in scale. Anthropic reported that from May to July, there were over 151 million interactions, with a peak of nearly 3 million interactions per day. The main focus was on extracting the complete inference process of Opus 4.6 and 4.7 for training Qwen 3.5, 3.6, and 3.7. 2. Dark side of the month: directly give Claude to Kimi users for answering. Nearly 300000 real user requests were forwarded to Claude within 10 days, some of which were accompanied by internal company codes and valid credentials; From May to July, there were over 23 million related interactions. 3. DeepSeek: It will also secretly forward user requests to Claude and specifically identify users who use tools such as Claude Code and OpenCode, and then select some requests to forward to Opus. Over 12.1 million times in 14 days. 4. Zhipu: specializes in extracting and cleaning Claude's reasoning process, and also uses American models to practice network security capabilities. Anthropic stated that Zhipu once attempted to distill Fable, but found that the protection was too strong and turned to Opus 4.6 and another American company's model. Over 3.4 million interactions in 17 days. 5. Xiaomi: Feed the real conversations and programming records generated by MiMo users back to Claude to generate SFT and reinforcement learning data. Over 400000 requests in 20 days; Anthropic even suspects that the free trial period of MiMo-V2-Pro may be used to attract more overseas developers to produce training materials. 6. SenseTime: Directly purchase chat records between users and Claude from third-party data providers. These data come from proxy and intermediary services, and some users are unaware that their conversations will be saved and sold. 7. MiniMax: Accused of providing intermediary services through a seemingly unrelated company's operational model, which only offers Anthropic and OpenAI models, without even MiniMax's own model. Anthropic believes that this service may be designed to collect user interactions from American models for training purposes. The above are all Anthropic's investigations, attributions, and allegations, which have not been independently verified at present. [Original link]