Anthropic exposes 7 companies in China for distillation techniques targeting large models 1. Alibaba: The largest scale. Anthropic claims that between May and July, there were over 151 million interactions, peaking at nearly 3 million per day. They primarily extracted the full reasoning process of Opus 4.6 and 4.7 to train Qwen 3.5, 3.6, and 3.7. 2. Dark Side of the Moon: Directly used Claude to answer Kimi users' queries. Nearly 300,000 real user requests were forwarded to Claude within 10 days, some even containing internal company codes and valid credentials. Between May and July, related interactions exceeded 23 million. 3. DeepSeek: Also secretly forwarded user requests to Claude, specifically identifying users who utilized tools like Claude Code and OpenCode, then selectively forwarding some requests to Opus. Over 12.1 million interactions in 14 days. 4. Zhipu: Specialized in extracting and cleaning Claude's reasoning process, and even used American models to train cybersecurity capabilities. Anthropic claims Zhipu once attempted to distill Fable but switched to Opus 4.6 and another American company's model after encountering strong defenses. Over 3.4 million interactions in 17 days. 5. Xiaomi: Re-fed real conversations and programming records from MiMo users back into Claude to generate SFT and reinforcement learning data. Over 400,000 requests in 20 days. Anthropic even suspects that the free trial period of MiMo-V2-Pro might have been used to attract more overseas developers to generate training material. 6. SenseTime: Directly purchased user chat records with Claude from third-party data vendors. These records came from proxy and relay services, and some users were unaware their conversations were being saved and sold. 7. MiniMax: Allegedly operated a model relay service through a seemingly unrelated company, offering only Anthropic and OpenAI models—without even including MiniMax's own models. Anthropic suspects this service was designed to collect user interactions with American models for training purposes. No wonder AI in China is advancing so fast