Chinese military labs distill OpenAI and Anthropic models into local defense systems
A Reuters and Jamestown Foundation review of more than 80 Chinese papers and patents found PLA-linked researchers distilling US model outputs into small domestic systems — PLA Unit 96941 used GPT-3.5 to summarize sensitive military source code, then trained a local model on those summaries to run inside classified networks. Distillation routes around chip export controls because what crosses the border is model outputs, not hardware, and Anthropic says distilled models can shed the safety safeguards of the originals. The findings land ahead of US-China talks on AI governance, where Washington has framed unauthorized extraction as an export-control and IP issue; OpenAI did not respond to requests for comment.
Source: defensenews.com ↗
Teaching a model the right answer is one thing but teaching it the reasoning behind the answer is much harder.
Why this matters
- → PLA researchers bypass chip sanctions by distilling US AI into local military systems
- → Distilled models lose original safety guardrails, enabling capabilities beyond vendor control
- → Affordable edge deployment — drones, submarines, satellites — with adversary-controlled reasoning