Three US security agencies have accused six Chinese AI developers of extracting billions of tokens from American frontier models through coordinated, industrial-scale distillation campaigns. China’s commerce ministry rejected the allegations Wednesday and said Beijing would take countermeasures if Washington used them to suppress Chinese companies.
The joint advisory from the National Security Agency, Cybersecurity and Infrastructure Security Agency and FBI names DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI. The agencies said the companies made millions of requests to variants of Anthropic’s Claude, OpenAI’s GPT, Google’s Gemini and xAI’s Grok from late 2024 onward, likely with Chinese government awareness.
Knowledge distillation is a standard training technique in which one model learns from another model’s outputs. The agencies said the conduct crossed into unauthorized extraction because the companies allegedly concealed their identities, breached provider terms and collected restricted capabilities at scale. Beijing called distillation a neutral technology used throughout the industry, including by US companies.
The advisory describes a distributed access system designed to survive provider blocks. It says Chinese developers used fraudulent accounts, third-party aggregators, remote cloud providers and gray-market API proxies known as transfer stations. Shared premium subscriptions reduced costs, while centralized routing systems shifted requests between access points and stripped identifying metadata.
The document assigns specific targets to individual labs. It says DeepSeek sought reasoning, agentic and specialized capabilities for its R1 and V3 models. Moonshot allegedly used Claude Fable 5 data to train Kimi K3 and GPT-4o data for Kimi K2. It also attributes model-development campaigns to Alibaba, MiniMax, StepFun and Z.AI. DeepSeek, Alibaba, Moonshot and Z.AI did not immediately respond to Associated Press requests for comment.
The allegations extend an April White House warning that treated covert distillation as a national-security issue. In July, the administration specifically accused Moonshot of using Anthropic outputs to build Kimi K3. The new document is a public technical advisory shared by three national security and law-enforcement agencies.
China’s commerce ministry said the United States was politicizing a common commercial practice to protect a monopoly in compute and AI. It also argued that Chinese open models are available to US developers. The ministry did not specify what countermeasures it would take. China’s foreign ministry separately urged Washington to stop making what it called unfounded accusations.
The advisory does not impose sanctions or export controls. It asks model providers, cloud platforms and API aggregators to detect unusual account and traffic patterns, coordinate threat intelligence and make targeted changes to responses served to suspected extraction campaigns.
Sources: National Security Agency, China Ministry of Commerce via Xinhua, Associated Press
–
By the Control Plane Editorial Team