According to CoinMeta, the National Security Agency of the United States, along with other authorities, issued a joint report accusing six Chinese companies—including DeepSeek, Moon’s Dark Side, Alibaba, Minimax, Step Star, and Zhipu–of bulk invoking the services of Claude, GPT, Gemini, and Grok since at least the end of 2024 to train their own models. This involved millions of requests and billions of token. The report indicates that DeepSeek was alleged to have used multiple Claude, GPT, Gemini, and Grok to generate training data for R1 and V3, while Moon’s Dark Side was accused of using Claude Fable data to train Kimi K3 and GPT-4O data to train Kimi K2. The U.S. side claims that these companies also bypassed regional restrictions through a large number of accounts, third-party API aggregators, and grey “transit stations,” and attempted to uncover the hidden inference processes of the models using certain prompts. Such behavior is referred to in the report as “malicious distillation.”