Anthropic says several China-based artificial intelligence companies attempted to extract advanced capabilities from its Claude models through large-scale unauthorized distillation campaigns.
In a new Threat Intelligence report, Anthropic said it identified and disrupted campaigns linked to seven AI labs based in China. The company named firms including Alibaba, Moonshot AI, DeepSeek, Zhipu — known internationally as Z.ai — and Xiaomi.
AI model distillation itself is a legitimate technique in which a powerful “teacher” model generates outputs that can be used to train a smaller model. Anthropic, however, describes these campaigns as “illicit distillation” because it says the companies covertly accessed Claude at industrial scale without authorization and attempted to replicate its capabilities.
According to Anthropic, the largest campaign was associated with Alibaba. The company says it observed more than 151 million Claude exchanges linked to the operation between May and July 2026, with activity reaching nearly three million interactions per day.
Anthropic alleges that the campaign targeted valuable Claude capabilities including reasoning, software engineering, agentic tasks, data analysis and long-horizon problem solving, with the resulting data used to improve Alibaba’s Qwen models.
The report also claims that Moonshot AI and DeepSeek sometimes routed requests from their own users through Claude without those users necessarily knowing their prompts were being sent to Anthropic.
Anthropic says it observed more than 23 million exchanges associated with Moonshot between May and July, while more than 12.1 million exchanges were attributed to DeepSeek during a 14-day period in July.
Some of those forwarded conversations allegedly contained sensitive corporate, personal and government-related information, raising additional privacy concerns.
Anthropic says it has responded by banning associated accounts, strengthening systems designed to detect model extraction, limiting access to detailed internal reasoning and introducing additional identity verification in regions where Claude is not officially supported.
The findings highlight a growing challenge for leading AI companies as competitors seek cheaper ways to reproduce the reasoning, coding and agentic capabilities of expensive frontier models.
The allegations remain Anthropic’s account of the activity. Reuters reported that the Chinese companies named in the report did not immediately respond to requests for comment.
