Anthropic says Chinese AI labs used Claude outputs to train models

Anthropic said it detected and disrupted unauthorized efforts by several Chinabased AI labs to use Claude exchanges for training their own models. The company said the activity involved Alibaba, Moonshot, DeepSeek and other firms, and that it appeared to rely on a technique known as distillation, where outputs from a stronger model are used to improve another system without permission. According to Anthropic’s report, the activity took place between December 2025 and August 2026 and included large volumes of requests routed through fraudulent accounts. The company said Alibabarelated activity was the largest it measured, with more than 151 million Claude exchanges over a threemonth period, while Moonshot and DeepSeek were also linked to extensive use of Claude responses and reasoning transcripts. Anthropic said some of the exchanges included sensitive information from users, companies and stateaffiliated actors, raising possible privacy and termsofservice concerns. Alibaba, Moonshot, DeepSeek, Xiaomi and Anthropic did not immediately respond to CNBC’s requests for comment.