CISA, NSA and FBI Warn of Chinese AI Firms’ Large-Scale Model Distillation Campaigns
Summary
A joint CISA, NSA and FBI advisory says China-based firms used millions of requests to extract capabilities from U.S. AI models, and recommends stronger verification, monitoring and output controls.
Key points
- The advisory names DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI as having extracted billions of tokens through millions of requests since at least late 2024.
- Agencies assess that the activity likely occurred with the Chinese government’s knowledge; this is an assessment, not a confirmed attribution.
- Reported tactics include API-proxy “transfer stations,” pools of premium accounts and centralized traffic routing to bypass restrictions, evade usage limits and obscure origins.
- Detection indicators include continuous high-volume usage, unusual subscription-to-API ratios, coordinated account pools and sudden changes in identifying metadata.
- Recommended safeguards include stronger identity checks, monitoring, response variation, rate limits, differential privacy and sharing indicators among model developers and service providers.
Article Details
- Event Type
- Joint cybersecurity advisory warning that China-based AI companies are conducting large-scale knowledge distillation campaigns against U.S. frontier AI models.
- Impact
- The agencies say the campaigns have extracted billions of tokens through millions of requests since at least late 2024, copying capabilities including reasoning, coding, and specialized functions. The activity may undermine safeguards and reduce the capability gap between U.S. frontier models and models developed by the named companies. The agencies assess that the activity likely occurred with the knowledge of the Chinese government.
People
Products
Claudeagencies say China-based AI companies have been running large-scale knowledge distillation campaigns to copy advanced capabilities from frontier models such as Claude, GPT, Gemini, and Grok.Claude Opus 4.8By mid-2026, Z.AI had distilled billions of tokens of data from GPT-5.5 and Claude Opus 4.8 to develop chain-of-thought reasoning capabilities.DeepSeek R1frontier AI models to help train DeepSeek R1 and V3, focusing on reasoning, writing, agentic functions, question answering and specialized tasks such as legal work.Geminiagencies say China-based AI companies have been running large-scale knowledge distillation campaigns to copy advanced capabilities from frontier models such as Claude, GPT, Gemini, and Grok.GPTagencies say China-based AI companies have been running large-scale knowledge distillation campaigns to copy advanced capabilities from frontier models such as Claude, GPT, Gemini, and Grok.GPT-5.5By mid-2026, Z.AI had distilled billions of tokens of data from GPT-5.5 and Claude Opus 4.8 to develop chain-of-thought reasoning capabilities.Grokagencies say China-based AI companies have been running large-scale knowledge distillation campaigns to copy advanced capabilities from frontier models such as Claude, GPT, Gemini, and Grok.Kimifrontier models to improve software engineering, mathematics, supervised fine-tuning and reinforcement learning capabilities in its Kimi models.