U.S. agencies have accused several Chinese artificial intelligence companies, including DeepSeek and Alibaba, of systematically extracting capabilities from leading American AI models.
According to U.S. officials, the activity took place on an industrial scale and was designed to reduce development costs while helping Chinese companies narrow the technology gap with U.S. AI leaders.
U.S. Agencies Issue Joint AI Security Warning
The Cybersecurity and Infrastructure Security Agency, National Security Agency, and Federal Bureau of Investigation issued the allegations in a joint advisory released on Tuesday.
The agencies said the campaigns had likely taken place with awareness from the Chinese government.
They also claimed that several Chinese AI companies had extracted billions of tokens across millions of interactions with U.S. models since at least late 2024.
The companies named in the advisory included DeepSeek, Alibaba, Moonshot AI, MiniMax, StepFun, and Z.AI.
DeepSeek and Alibaba Named in Model Extraction Claims
U.S. agencies said DeepSeek carried out organized campaigns targeting advanced reasoning capabilities and specialized AI functions.
According to the advisory, those efforts were linked to the development of DeepSeek’s R1 and V3 models.
Alibaba was also accused of using large-scale AI distillation techniques to improve its Qwen model family.
Model distillation generally involves using the outputs of a larger or more capable AI system to help train or improve another model.
U.S. Says Chinese Firms Targeted Major AI Models
The advisory said DeepSeek targeted several major U.S. and Western AI systems.
These allegedly included OpenAI’s GPT models, Anthropic’s Claude, Google’s Gemini, and xAI’s Grok.
According to U.S. officials, the companies sought to extract capabilities related to reasoning, coding, autonomous agent functions, and question-and-answer performance.
The agencies argued that this allowed developers to gain useful model capabilities without bearing the full cost of building frontier AI systems from the ground up.
Multiple Methods Allegedly Used to Bypass Restrictions
U.S. officials said the companies used several different methods to gain access to restricted AI services.
These included direct API access, cloud computing platforms, third-party aggregators, and proxy networks.
The advisory also referred to a gray market of intermediaries known as “transfer stations,” which were allegedly used to bypass geographic restrictions and other safeguards.
The agencies said fraudulent accounts, bulk subscriptions, and automated systems were also used in some cases to avoid detection.
AI Model Distillation Becomes a Growing Security Concern
The allegations highlight growing concern over AI model extraction and distillation as competition in artificial intelligence intensifies.
Training advanced AI models can require enormous amounts of computing power, data, and capital.
If a company can reproduce valuable capabilities by repeatedly querying an existing model, development costs can potentially be reduced significantly.
This has made model extraction an increasingly important issue for AI companies seeking to protect proprietary technology.
U.S. AI Companies Urged to Strengthen Monitoring
The U.S. agencies recommended that American AI companies improve their defenses against suspected model extraction campaigns.
They advised providers to strengthen monitoring for unusual or highly automated usage patterns.
Companies were also encouraged to modify model responses when suspicious distillation activity is detected.
In addition, the agencies called for greater intelligence sharing between AI model developers, cloud providers, and API aggregation platforms.
U.S.-China AI Competition Continues to Intensify
The accusations come as competition between the United States and China over advanced artificial intelligence continues to grow.
AI models, semiconductors, data centers, and computing infrastructure have become increasingly important areas of strategic competition between the two countries.
The latest advisory suggests that model security is also becoming a larger part of that rivalry.
As frontier AI systems become more capable and expensive to develop, protecting access to their underlying capabilities could become an increasingly important priority for major technology companies and governments.
