DeepSeek under fire: Anthropic accuses Chinese AI firm of misusing Claude for unauthorized model training

US artificial intelligence company Anthropic has accused Chinese AI firms of misusing Claude and siphoning data for AI model training purposes.

According to the tech firm, the companies including DeepSeek, MiniMax, and Moonshot AI used a technique called “distillation” in which these AI firms extract and leverage outputs from a more powerful AI system to rapidly boost the performance of low-scale models.

Taking to X, Anthropic posted, “These labs created over 24,000 fraudulent accounts and generated over 16 million exchanges with Claude, extracting its capabilities to train and improve their own models.”

Anthropic tracked 150,000 conversations from DeepSeek, aiming to improve alignment and foundational logic. Similarly, Moonshot AI was found to be involved in 3.4 million conversations to boost agentic reasoning power, coding analysis and improve computer vision.

“These campaigns are growing in intensity and sophistication,” the company said in a statement. “The window to act is narrow.”

Distillation is a common and legitimate practice within AI development and it is often used by companies to develop cheaper and smaller versions of their models.

The company further argued that such practice if employed illicitly can pose a serious threat to national security as the foreign labs can remove safeguards and feed the model capabilities with their own military, intelligence, and surveillance systems.

Given the intensity and frequency of cyberattacks, it is important to address them with rapid, coordinated action among industry players, policymakers, and the broader AI community.

Besides Anthropic, its long-held rival OpenAI also made similar accusations earlier this month, stating Chinese AI firms are using the distillation technique to capitalize on the advanced capabilities of ChatGPT models and other US frontier labs.

The accusation comes as DeepSeek is planning to release its new and advanced AI model as early as next week.