A major Chinese AI developer, Zhipu, has found itself at the center of a scandal. It has emerged that their latest model, GLM-5.2, was trained on the responses of Anthropic's flagship developments — Claude, as well as GPT from OpenAI. This statement was made at the Aspen Security Forum by Tarun Chhabra, head of national security policy at Anthropic.
In my assessment, this incident is not just an isolated case but part of a systemic problem. Chhabra directly stated that distillation, i.e., extracting knowledge from more powerful models through their responses, is practiced by "virtually all" leading Chinese labs. This applies not only to Claude but also to other American developments. Anthropic is forced to block up to several million suspicious accounts weekly in an attempt to stem this flow.
The expert's main conclusion: the US maintains its lead in advanced AI models, ahead of China by six to nine months. However, without this illegal distillation, the gap would be significantly larger — up to a year and a half. He rightly called this behavior by Beijing "hostile," emphasizing that competitors are extracting the most valuable intellectual property.
New Evidence and the Scale of the Threat
Separately, Chhabra presented fresh data on Zhipu AI. Research company Graphistry conducted an analysis: by Cohen's kappa coefficient, GLM-5.2's responses matched GPT-5.5 at 0.8 and Opus 4.8 at 0.76. For comparison, between GPT-5.5 and Opus 4.8 themselves, this figure was only 0.63. This is direct evidence that the model was "trained" on others' results. Zhipu AI has yet to confirm or deny this data. DeepSeek is also continuing a similar "hostile campaign," confirming the systemic nature of the problem.
Against this backdrop, Washington is tightening its policy. Earlier, the White House launched the Gold Eagle initiative — a platform for sharing vulnerability data between leading US AI companies and government agencies. Treasury Secretary Scott Bessent stated that the regulator will use advanced models to stay ahead of adversaries. Recall that the Trump administration previously blocked foreign access to Anthropic's most powerful models — Fable 5 and Mythos 5, leading to their temporary shutdown for all users.
My comment: This situation is a vivid example of how geopolitical competition is turning into a technological war. Distillation, while a known method, in this context becomes a tool of industrial espionage. The AI market is entering a phase where intellectual property becomes the main battlefield, and companies that fail to protect their algorithms risk losing multi-billion-dollar development investments. This threat can no longer be ignored.