US Agencies Accuse Chinese AI Firms of Industrial Scale Model Distillation

The National Security Agency, Cybersecurity and Infrastructure Security Agency, and FBI accused six China based AI companies of extracting capabilities from US frontier models at industrial scale in a joint cybersecurity advisory released September 8. The agencies named DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI.

The advisory alleges that the companies extracted billions of tokens through millions of requests since at least late 2024. It says they targeted variants of Claude, GPT, Gemini, and Grok to obtain reasoning, coding, software engineering, agent functions, and other capabilities for training their own models.

According to the agencies, the companies routed requests through native APIs, cloud providers, third party aggregators, fraudulent accounts, and proxy services called transfer stations. The campaigns allegedly used bulk premium subscriptions, automated switching between access routes, prompt injection, and systems that assessed whether providers had altered their responses.

The agencies recommend monitoring unusual account activity, immediate maximum usage from new subscriptions, sustained activity without normal idle periods, and coordinated requests across multiple services. They also recommend altering responses to confirmed distillation campaigns, limiting query volume, strengthening account verification, and sharing infrastructure and behavioral indicators among model providers, cloud platforms, and API aggregators.