U.S. Accuses Chinese AI Firms of Extracting Model Capabilities

U.S. security agencies have accused several major Chinese artificial intelligence companies, including DeepSeek and Alibaba, of systematically extracting capabilities from advanced U.S. AI models to reduce development costs and close the technology gap.

The Cybersecurity and Infrastructure Security Agency (CISA), National Security Agency (NSA) and Federal Bureau of Investigation (FBI) said in a joint advisory released Tuesday that the activity was carried out on an industrial scale and likely occurred with the awareness of the Chinese government.

According to the advisory, DeepSeek, Alibaba, Moonshot AI, MiniMax, StepFun and Z.AI have extracted billions of tokens through millions of interactions with U.S. AI models since at least late 2024.

U.S. officials said DeepSeek conducted coordinated campaigns aimed at obtaining advanced reasoning capabilities and specialized functions that could support development of its R1 and V3 AI models. Alibaba was similarly accused of using large-scale model distillation techniques to enhance its Qwen AI family.

The companies allegedly accessed U.S. models through several channels, including application programming interfaces (APIs), cloud service providers, third-party aggregators and a gray market of proxy services known as “transfer stations.” The advisory said fraudulent accounts, bulk subscriptions and automated systems were also used to circumvent geographic restrictions and security controls.

DeepSeek specifically targeted leading AI models including OpenAI's GPT, Anthropic's Claude, Google's Gemini and xAI's Grok, according to the agencies. The campaigns reportedly focused on extracting capabilities involving advanced reasoning, computer coding, AI agents and question-and-answer optimization.

The allegations highlight growing U.S.-China competition over artificial intelligence as Washington seeks to protect cutting-edge AI technology from overseas rivals.

CISA, the NSA and FBI urged U.S. AI developers to improve detection of unusual usage patterns that could indicate model distillation attempts. The agencies also recommended modifying model responses when suspicious activity is detected and increasing intelligence sharing among AI developers, cloud providers and API aggregators to limit future capability extraction.