Anthropic claims three Chinese AI laboratories stole their data a few days after OpenAI issued its warning

US artificial intelligence firm Anthropic has alleged that three Chinese AI labs — DeepSeek, Moonshot AI and MiniMax — extracted the capabilities of its flagship Claude model to enhance their own systems.

Claude, Anthropic’s AI assistant, is designed to perform a wide range of tasks, including coding, mathematics, reasoning, research and general conversation. According to the San Francisco-headquartered company, the three labs generated more than 16 million interactions with Claude using approximately 24,000 fraudulent accounts, in violation of its terms of service and regional restrictions.

The claims come days after rival OpenAI said that DeepSeek and other Chinese firms were extracting large volumes of data from ChatGPT. OpenAI had warned that some startups were employing advanced “distillation” techniques to replicate the technologies developed by American AI companies.

In a blog post published Monday, Anthropic said the Chinese labs relied on a technique known as “distillation,” which involves training a smaller or less capable model on the outputs of a more advanced one. While distillation is a common industry practice to make models cheaper and faster to run, Anthropic alleged that in this case it was used to copy Claude’s most advanced capabilities — including reasoning, coding and tool use — without authorisation.

The company said it linked the activity to specific labs through IP tracking, metadata analysis, infrastructure patterns and coordination signals. In certain instances, Anthropic added, industry partners reported observing similar behaviour on their own platforms.

Anthropic detailed the scale of activity attributed to each lab. DeepSeek allegedly conducted more than 150,000 exchanges, focusing on reasoning tasks and extracting step-by-step internal explanations. Moonshot AI is said to have carried out over 3.4 million exchanges, targeting agentic reasoning, coding and computer-use systems. MiniMax reportedly generated more than 13 million exchanges, concentrating on coding and orchestration tools, and adjusted its activity within 24 hours of a new Claude model release.

Anthropic warned that models built through unauthorised distillation could lack critical safety safeguards, including protections against misuse in areas such as bioweapons research, cyberattacks and mass surveillance. It also argued that such extraction campaigns risk undermining US export controls by enabling foreign labs to replicate advanced American AI systems without independently developing them.

The company, led by chief executive Dario Amodei, said it is strengthening its defences by enhancing detection systems for suspicious usage patterns and coordinated account activity. It is also tightening account verification processes, particularly for high-risk access pathways.

The allegations add to mounting tensions between US and Chinese AI developers, as competition intensifies over advanced models and the infrastructure that powers them.