Anthropic once again targets Chinese large models, exposing the distillation methods of 7 companies.

Beating AI News Flash: Anthropic's latest report states that over the past few months, it has discovered and banned Claude distillation activities from 7 Chinese AI companies. Alibaba, Moonshot AI, DeepSeek, Zhipu, Xiaomi, SenseTime, and MiniMax were all named. Distillation itself is a common training method; Anthropic's allegations focus on large-scale circumvention of restrictions, scraping Claude outputs and user data.

1. Alibaba: the largest scale. Anthropic says that from May to July, there were over 151 million interactions, with a peak approaching 3 million per day, mainly extracting the complete reasoning processes of Opus 4.6 and 4.7 for training Qwen 3.5, 3.6, and 3.7.

2. Moonshot AI: directly used Claude to answer on behalf of Kimi users. Within 10 days, nearly 300,000 real user requests were forwarded to Claude, some of which also carried internal enterprise code and valid credentials; from May to July, related interactions exceeded 23 million.

3. DeepSeek: also secretly forwarded user requests to Claude, and specifically identified users of tools such as Claude Code and OpenCode, then selected some requests to forward to Opus. Over 14 days, more than 12.1 million times.

4. Zhipu: specifically extracted and cleaned Claude's reasoning processes, and also used U.S. models to train cybersecurity capabilities. Anthropic says Zhipu at one point tried to distill Fable, and after finding the protections too strong, shifted to Opus 4.6 and another U.S. company's model. Over 17 days, more than 3.4 million interactions.

5. Xiaomi: fed real conversations and programming records already generated by MiMo users back into Claude to generate SFT and reinforcement learning data. Over 20 days, more than 400,000 requests; Anthropic even suspects that MiMo-V2-Pro's free trial period may also have been used to attract more overseas developers to generate training material.

6. SenseTime: directly purchased chat records between users and Claude from third-party data vendors. This data came from agents and relay services, and some users did not know their conversations would be saved and sold.

7. MiniMax: accused of operating a model relay service through a company with no apparent connection on the surface, and it only offered Anthropic and OpenAI models, not even MiniMax's own models. Anthropic judged that this service may have been intended to collect user interactions with U.S. models for training.

The above are all Anthropic's investigations, attributions, and allegations, and have not been independently verified.