Researchers working for China's military used outputs from OpenAI and Anthropic models to help train their own defense systems, according to a report from Plataforma Media circulated in Google News' defense AI feed.
The report frames the story around the method itself, explaining how the technique works. The broad approach it describes — feeding one AI system's answers into the training of another — is the industry practice usually called distillation. In plain terms, instead of building a powerful model from scratch at enormous cost, you query someone else's finished model, collect its responses, and use that harvested material to teach a smaller or newer system to behave similarly. The student never sees the teacher's underlying code or weights; it learns from the answers alone.
Beyond that, hard details are thin. The available source is a single summary, and it does not specify which models were queried, which Chinese institutions were involved, when the work took place, or how effective the resulting defense systems are. Neither OpenAI nor Anthropic is quoted in the material at hand, and no independent confirmation appears in it.
Still, the claim lands on a sensitive nerve. Both companies' usage policies and U.S. export controls are built around restricting who gets access to frontier AI capability — but distillation routes around hardware bans and licensing agreements entirely, since anyone with an API key and patience can extract behavior from a model rather than steal the model. Chip controls cannot stop a text conversation.
It matters because it suggests the most valuable thing American AI labs produce may be far harder to keep inside the country than the chips everyone has been arguing about.