Executive Summary
China-based artificial intelligence companies, including DeepSeek, Alibaba, Moonshot AI, MiniMax, StepFun, and Z.AI, are executing systematic extraction of proprietary functionalities from U.S. frontier AI models. This industrial-scale knowledge distillation strategy is used to bridge technological gaps, bypass regional restrictions, and significantly reduce development costs and timelines. The campaigns target reasoning capabilities, coding, and specialized agentic functions using sophisticated evasion techniques.
Technically, these actors utilize a gray market of API proxies known as 'transfer stations' to bypass geographic restrictions and evade provider safeguards. They employ automated request metadata sanitization and centralized routing infrastructure to obfuscate user identity and maintain high-volume query throughput. Notable targets include models from the GPT-4/5 series, Claude 3/4/5 families, and Gemini 2 series.
This activity represents a direct threat to U.S. technological leadership and intellectual property. By extracting chain-of-thought (CoT) reasoning and reward model functions, these companies are effectively cloning the behavior and performance of advanced American LLMs. The scale of these operations, involving millions of queries, indicates that distillation is the critical core of their model development lifecycle.
Key Details
Threat Name
China-Based Industrial-Scale AI Distillation
Affects
—
Adversary
DeepSeek Other Adversaries and Aliases: Moonshot AI; Alibaba; MiniMax; StepFun; Z.AI
MITRE Techniques
—
Malware/Tools
None identified