Chinese AI developers siphon US frontier models at industrial scale using shadow proxy networks
- Leading Chinese AI startups operated massive proxy networks routing synthetic reasoning queries through Singapore, Japan, and U.S. consumer IPs.
- Harvested millions of reasoning trajectories from Claude and GPT-4 to distill into domestic models marketed as indigenous breakthroughs.
- Circumvents U.S. cloud compute export controls and terms of service through covert automated scraping.
China’s AI champions are running distillation pipelines on Western servers. A joint intelligence advisory from the NSA, FBI, and CISA, reported by The Register, revealed that firms including DeepSeek and Alibaba Cloud siphoned reasoning data from American frontier models at an industrial scale. The companies routed millions of automated distillation queries through shadow commercial proxies and gray-market relay networks, harvesting advanced coding and mathematical logic from GPT, Claude, and Gemini to train domestic architectures.
The pipeline is a coordinated campaign of state-backed model distillation and export control evasion, wrapped in propaganda disinformation. It recalls decades of illicit technology transfer, from the 2006 Hanxin chip scandal to aerospace blueprint theft for the C919. Cyberspace Administration of China propaganda loudly proclaims pure “indigenous innovation.” In reality, domestic tech giants secretly tap foreign research to mask foundational weaknesses at home.
This intellectual property theft directly fuels Beijing’s military AI ambitions. Harvested reasoning models are adapted for autonomous drone swarms, intelligence processing, and cyber warfare. By laundering foreign code through private champions, the party-state turns Western innovations into instruments of authoritarian power.
Chinese AI developers on developer forum Juejin openly discussed methods to bypass OpenAI and Anthropic geo-blocks and rate limits using residential proxy farms.
State media vigorously claimed all leading domestic LLMs are 'fully autonomous and sovereign', silencing technical benchmarks showing model weights distilled from Western frontier models.
Proxy networks utilized tens of thousands of hijacked consumer IoT devices across Southeast Asia to disguise synthetic data extraction queries.