US says Chinese firms extracted billions of tokens from frontier AI models

U.S. cybersecurity agencies warn that six China-based AI firms (DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, Z.AI) have performed industrial-scale model distillation attacks against frontier models from Anthropic, OpenAI, Google, and xAI, extracting billions of tokens via millions of distributed API requests. The advisory attributes high sophistication and likely state awareness to the operations, details tactics such as chain-of-thought extraction and automated failovers, lists behavioral and infrastructure indicators for detection, and recommends modifying responses and sharing intelligence to mitigate these campaigns.