Anthropic alleges Chinese AI labs used 190M Claude exchanges for training

Anthropic has released a threat intelligence report alleging that seven Chinese AI laboratories have been using its Claude models to train their own artificial intelligence systems through a process known as distillation.

The report claims these companies used approximately 190 million exchanges with Claude to improve their models. Anthropic identified the participating labs as Alibaba, Moonshot AI, DeepSeek, Zhipu, Xiaomi, SenseTime, and MiniMax. The company alleges that these entities used fraudulent accounts, residential proxies, and virtual payment cards to bypass restrictions and access the models.

Alibaba is cited as having conducted the largest campaign, allegedly generating over 151 million exchanges between May and July 2026. These interactions reportedly targeted Claude Opus 4.6 and 4.7 reasoning traces to improve Alibaba’s Qwen 3.5, 3.6, and 3.7 models. Anthropic also noted that Moonshot AI and DeepSeek allegedly forwarded actual user conversations to Claude to facilitate training.