US agencies allege Chinese AI firms ran industrial-scale model distillation campaigns

The NSA, CISA and FBI have accused DeepSeek, Moonshot AI and other Chinese artificial-intelligence companies of conducting “industrial-scale” campaigns to extract data and capabilities from American frontier models.

According to the joint cybersecurity advisory, DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI allegedly obtained billions of tokens through millions of exchanges or requests since 2024. The models identified as targets include Anthropic’s Claude, OpenAI’s GPT, Google’s Gemini and xAI’s Grok.

The agencies said DeepSeek used data and capabilities from Claude, Gemini, GPT and Grok models to train its own systems, including its R1 reasoning model. They also alleged that Moonshot AI extracted significant data from Claude’s Fable to train Kimi K3, while Kimi K2 used data from GPT-4o.

Model distillation involves using the output of a more powerful model to train another model. The advisory sets out mitigation measures for US companies seeking to counter what the agencies describe as malicious distillation campaigns.

The allegations follow similar claims from OpenAI and Anthropic involving DeepSeek, Moonshot AI and MiniMax. OpenAI and Microsoft have previously said they banned accounts suspected of distilling OpenAI technologies.

The advisory’s evidence and methodology were not detailed in the supplied material, which contains no response from the named Chinese companies or independent verification. Engadget reported the release of the joint advisory on Wednesday.