Chinese AI Firms Accused of Extracting US Model Capabilities
US agencies reveal industrial-scale AI knowledge distillation campaigns.
1US Agencies Expose Chinese AI Extraction
The NSA, CISA, and FBI have accused six Chinese AI firms—DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI—of conducting industrial-scale extraction of US AI models. Since late 2024, these companies allegedly sent millions of requests to models like Claude, GPT, Gemini, and Grok, extracting billions of tokens. This activity is described as central to their AI development strategy, not just a supplementary tactic.
2Targeted US AI Models
The Chinese firms reportedly targeted advanced US AI models, including Claude, GPT, Gemini, and Grok, to extract proprietary functionalities. These models contain capabilities like reasoning, coding, and agentic skills, which took years and significant computing power to develop. The extraction campaigns aimed to replicate these features to accelerate China’s AI development.
3DeepSeek’s Organized Campaign
DeepSeek allegedly ran a systematic campaign against multiple versions of Claude, GPT, and Gemini between late 2024 and mid-2025. The extracted data included specialized knowledge, such as legal expertise, and chain-of-thought reasoning. This effort reportedly helped develop DeepSeek’s R1 and V3 models, enhancing their capabilities in areas like agentic functioning and question-answer optimization.
4Moonshot AI’s Extensive Extraction
Moonshot AI is accused of distilling 18 different US models, including Anthropic’s Fable 5, to train its Kimi-K2 and Kimi-K3 models. The company used millions of queries to extract capabilities in agentic reasoning, coding, data analysis, and visual processing. This sophisticated approach allowed Moonshot AI to significantly shorten its development timelines.
5Sophisticated Tactics to Avoid Detection
Chinese AI firms employed advanced tactics to evade detection, such as spreading requests across multiple accounts, models, and platforms. They used native APIs, remote cloud providers, and third-party aggregators to obscure user metadata. Proxies and gray tech markets were also leveraged to bypass geographic restrictions and terms of use.
6Industrial-Scale Knowledge Distillation
The extraction campaigns involved industrial-scale knowledge distillation, a technique where smaller models are trained to replicate larger ones. While distillation is a legitimate practice, the US agencies argue that the Chinese firms went beyond acceptable limits, systematically extracting proprietary functionalities to accelerate their AI development.
7Impact on US AI Development
The extraction campaigns pose a significant threat to US AI innovation. By replicating advanced capabilities developed over years, Chinese firms can reduce their development timelines and costs. This undermines the competitive edge of US AI companies and raises concerns about intellectual property theft on a massive scale.
8Broader Implications for AI Competition
The allegations highlight the intensifying competition in AI between the US and China. If proven, these activities could lead to stricter regulations and safeguards for AI models. The incident underscores the need for international cooperation to address intellectual property theft and ensure fair competition in the AI sector.
Original article: Security Affairs — Chinese AI Firms Accused of Extracting US Model Capabilities