Artificial intelligence firm Anthropic has detected and disrupted large-scale attempts by Chinese AI laboratories to steal its technology through a process known as "distillation." These illicit operations, described by Anthropic as the largest yet, involved companies including Alibaba, Moonshot, and DeepSeek using outputs from Anthropic's advanced Claude models to train their own less capable AI systems without permission.
Anthropic announced on Thursday that it identified the practice where a less-advanced AI is trained on the responses of a more sophisticated one to mimic its capabilities. This approach allows companies to rapidly improve their AI models without the extensive research and development typically required.
In one significant incident over a 10-day period this year, Moonshot, based in Beijing, allegedly rerouted 300,000 of its customer inquiries to Claude. The company then presented these responses to its own customers without disclosing that they were generated by Anthropic's AI. To facilitate this, Moonshot reportedly used over 5,000 fake accounts based in Singapore and Japan, collecting the exchanged data for training purposes.
Alibaba's alleged distillation operation was identified as the largest, involving more than 151 million interactions with Claude between May and July. Anthropic stated that entities linked to Alibaba used Claude's generated responses to train their Qwen models and to enhance their model architecture. This activity, which reached nearly 3 million exchanges daily and involved over 3,500 fraudulent accounts, also included sensitive information from individual users, multinational corporations, and state-affiliated entities, raising concerns about privacy law and terms of service violations.
DeepSeek, another Chinese AI company known for its cost-effective models, is accused of employing similar tactics. Anthropic detected over 12 million distillation attacks by DeepSeek within a two-week span in July. Like Moonshot, DeepSeek reportedly redirected user exchanges to Claude without informing its customers.
These incidents represent a growing concern within the U.S. AI industry, as leading companies race to develop cutting-edge AI systems. The attempts by less advanced Chinese AI firms to replicate proprietary technology underscore the competitive landscape and potential risks associated with the rapid advancement of artificial intelligence.
Anthropic's threat intelligence report also noted other disruptive activities by major Chinese AI companies between December 2025 and August 2026, including cyber operations, surveillance, scams, fraud, and misuse related to biological and conventional weapons development. Alibaba, Moonshot, and DeepSeek did not immediately respond to requests for comment.