Anthropic has publicly named seven Chinese AI companies and models that it says were involved in large-scale efforts to extract data from Claude for model training. The disclosure has renewed attention on AI model distillation, data security and the increasingly competitive global AI market.
Anthropic Reports Large-Scale Model Distillation
According to Anthropic, several Chinese AI companies allegedly used large numbers of accounts and automated requests to interact with Claude and collect its responses.
The companies and models named in the report include Alibaba’s Qwen, Moonshot AI’s Kimi, DeepSeek, Z.ai, Xiaomi, SenseTime and MiniMax.
Anthropic said the activities involved systematic interactions with Claude, particularly in areas such as reasoning and coding, and were designed to obtain large amounts of model-generated data.
Qwen Activity Reached a Significant Scale
Anthropic said Alibaba’s Qwen was associated with one of the largest activity clusters identified in its investigation.
Between May and July, the company said the activity involved approximately 151 million interactions. Anthropic also reported peak daily request volumes of around three million.
The company alleged that more than 3,500 accounts were used to collect responses from Claude models, including Opus 4.6 and 4.7, with the resulting data potentially being used to improve or train Qwen models.
Other Chinese AI Models Also Named
Anthropic also linked Kimi to activity involving the transfer of user requests to Claude and the subsequent return of generated results.
The company said some of the data involved potentially sensitive information, including content related to military facilities, government systems and corporate infrastructure.
DeepSeek was also accused of using cross-session techniques to reproduce Claude’s reasoning patterns and redirect developer traffic toward Claude-related tools such as Claude Code and its SDK.
Z.ai was reportedly involved in attempts to evaluate other US AI models after efforts involving its own Fable model did not produce the expected results.
Distillation Becomes a New AI Industry Dispute
Model distillation is a common AI development technique in which outputs from a more capable model are used to help train or improve another model.
Anthropic argues that the activities it identified went beyond ordinary model use and involved large-scale automated extraction designed to reproduce capabilities developed by Claude.
The allegations highlight a broader debate over how AI companies can use third-party models, training data and generated outputs. As competition between AI developers intensifies, questions surrounding intellectual property, data security and acceptable model usage are becoming increasingly important.