U.S. Agencies Accuse Chinese AI Firms of Systematic Model Distillation

Summary

The NSA, CISA and FBI accuse Chinese AI companies of conducting a deliberate, industrial-scale effort to extract capabilities from U.S. frontier models through knowledge distillation. A joint advisory says the activity has occurred since at least late 2024 and involved billions of tokens across millions of exchanges with models including Anthropic’s Claude, OpenAI’s ChatGPT, Google Gemini and xAI’s Grok. The advisory names DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI. U.S. agencies allege that DeepSeek used outputs from multiple Claude, Gemini, ChatGPT and Grok versions to create synthetic training data for its R1 and R3 models, while Moonshot allegedly queried 18 U.S. models to develop Kimi-K2 and Kimi K3 capabilities in areas such as agentic reasoning, coding, data analysis and computer vision. The companies are accused of routing requests through different accounts, models, platforms, native APIs, cloud providers, aggregators and proxies to obscure metadata and bypass geographic restrictions, terms of use and safeguards. The advisory says the activity is encouraged but not directly ordered by political leaders in Beijing. It calls for coordinated information-sharing among governments, industry and allied countries, while acknowledging that legitimate model distillation and weight sharing also occur in AI research and development. The allegations are presented by U.S. agencies and are not independently established in the article.