动察 Beating AI Flash: The U.S. National Security Agency (NSA), FBI, and CISA issued a joint report naming DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Zhipu AI. U.S. authorities allege that since at least late 2024, these companies have bulk-harvested outputs from Claude, GPT, Gemini, and Grok to train their own models, involving millions of requests and tens of billions of tokens cumulatively.

The report also directly lists specific models. DeepSeek is accused of using multiple Claude, GPT, Gemini, and Grok models to generate training data for R1 and V3. Moonshot AI is accused of using Claude Fable 5 data to train Kimi K3 and GPT-4o data to train Kimi K2.

U.S. authorities claim these companies also bypassed regional restrictions through numerous accounts, third-party API aggregators, and gray-market "relay stations," and used prompts to attempt extracting hidden reasoning processes from the models. The report therefore labels such behavior as "malicious distillation." Distillation itself is a common model training method; the U.S. action targets unauthorized large-scale calls and circumvention of access restrictions.