Adversary Intelligence — 2026-07-31

Chinese Military Researchers Use OpenAI and Anthropic Models to Train Defense Systems via Distillation

BLUFCurrent export controls targeting chips and direct API access leave model outputs unregulated, allowing PLA-linked institutions to harvest frontier AI reasoning through distillation faster than Washington can close the gap.

A Reuters review of more than 80 Chinese academic papers and patents, incorporating research compiled by the Jamestown Foundation, found that researchers linked to the People's Liberation Army and other Chinese military institutions have used outputs from OpenAI and Anthropic models to train domestic defense-related AI systems through model distillation 12. Reuters identified specific cases: PLA Unit 96941 used GPT-3.5 to summarize sensitive military source code for a domestic model run entirely on Chinese military networks, and China's National University of Defense Technology and Academy of Military Sciences used distillation for image-processing on drones and target recognition in simulated maritime operations 1. Researchers at the North University of China, which the report describes as linked to the weapons industry, used Anthropic's Claude 3 Haiku to generate synthetic training data for social media content-moderation models 13. Anthropic told Reuters it does not provide commercial access to Claude in China or to Beijing-controlled firms and said distilled models can lose the original systems' safety safeguards 1; the White House, Pentagon, Chinese foreign ministry, PLA, and OpenAI did not respond to Reuters' requests for comment 1.

Analysis
Commercial access restrictions did not stop capability transfer: PLA-linked institutions obtained frontier reasoning through model outputs rather than direct API access, sidestepping safeguards built for military end-use. Distillation strips the teacher models' safety alignment, so derivative systems for content moderation, drone targeting, and maritime target recognition carry none of the original guardrails, exposing a gap in export-control architecture that restricts chips and direct access but not reasoning traces embedded in outputs. Sourcing rests on a single exclusive Reuters investigation built atop independent Jamestown Foundation document research, with other outlets reproducing that account nearly verbatim rather than corroborating it independently; the papers reviewed were openly published in Chinese academic and military journals rather than concealed, consistent with routine research documentation rather than a clandestine extraction program. The finding lands days before US-China AI governance talks, giving Washington a concrete case for pressing output-level restrictions rather than access-level ones alone.

5 sources

Exclusive-Chinese military researchers tap US AI models to train defence systems - Reuters
Chinese Research Details Distillation for Military Use - Jamestown Foundation
Chinese military used OpenAI, Anthropic AI outputs to train their defence systems: Report - ChinaTechNews.com
Chinese military researchers tap US AI models to train defense systems - Defense News
China taps US AI models to train defense systems - Taipei Times