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

Summary

A joint U.S. cybersecurity advisory accuses Chinese AI firms of using millions of exchanges with U.S. frontier models since late 2024 to gather training data, with proxies and other tactics allegedly used to evade detection.

Key points

The NSA, CISA and FBI advisory names DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI.
The agencies say the firms used billions of tokens across millions of exchanges with models including Claude, ChatGPT, Gemini and Grok.
DeepSeek allegedly used generated data from U.S. models to train R1 and R3; Moonshot AI allegedly used millions of queries to train Kimi-K2 and Kimi K3.
The advisory says firms spread requests across accounts, models and platforms and used APIs, cloud providers, third-party aggregators and proxies to obscure activity or bypass restrictions.
The agencies call for coordinated information-sharing among the U.S. government, private industry and allied nations.
The article notes that distillation can also be used legitimately; the agencies characterize the alleged activity as targeted and industrial-scale.

Article Details

Event Type: U.S. government cybersecurity advisory alleging systematic, industrial-scale distillation of U.S. AI models by Chinese AI companies.
Impact: The agencies allege that the companies used billions of tokens and millions of exchanges with U.S. frontier models to extract capabilities and train their own models, while using obfuscation to avoid detection. The article reports no specific service disruption or direct operational damage; the claims are attributed to the U.S. agencies.