U.S. accuses china of systematic AI model distillation

U.S. federal agencies have accused Chinese AI firms of engaging in a systematic and industrial-scale effort to illegally distill U.S. frontier AI models. A joint advisory from the NSA, CISA, and FBI highlights that since 2024, Chinese companies like DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI have spent billions of tokens interacting with U.S. models such as Claude, ChatGPT, Gemini, and Grok. These companies allegedly used the data to enhance their own models, including DeepSeek’s R1 and R3 models and Moonshot AI’s Kimi-K2 and Kimi K3 models. The advisory notes that Chinese firms employ sophisticated tactics to avoid detection, such as routing requests through multiple accounts and platforms. In June, White House official Michael Kratsios accused Moonshot AI of distilling Fable 5 to train its models. The agencies emphasize that while sharing weights is common in AI research, Chinese companies are accused of conducting aggressive, malicious distillation at an industrial scale.