Anthropic accuses Chinese AI labs of using thousands of fake accounts to access its models

Anthropic says an illicit ecosystem is using fake identities, stolen bank cards and dark-web infrastructure to get around restrictions on advanced American AI models.

The company says the activity includes model distillation, a technique that uses one AI model’s answers to train another. Distillation can be legal with the technology owner’s permission, but Anthropic alleges that some companies are trying to evade its controls and illegally reproduce the capabilities of Claude.

“

There is an entire illicit ecosystem trying to obtain access to Claude and other models

“There is an entire illicit ecosystem trying to obtain access to Claude and other models,” said Jacob Klein, Anthropic’s head of threat intelligence. He said the networks use “any means necessary” to bypass security systems and create accounts on a very large scale.

“

Any means necessary

Anthropic pointed to several Chinese AI companies, including Moonshot AI. Klein alleged that the development of Moonshot’s Kimi K3 involved unauthorized use of the latest Claude model. The company said earlier this year that Moonshot, DeepSeek and MiniMax had used distillation techniques on its models. It also accused Alibaba, which develops the Qwen family, of carrying out a large-scale operation to acquire Claude’s capabilities. None of the companies responded to CNBC’s requests for comment.

The alleged activity is not limited to China, cybersecurity experts said. Similar efforts have been associated with actors in Iran, Russia and North Korea, where American companies restrict access to services including Claude, Gemini and ChatGPT. OpenAI and Google have also reported comparable attempts.

“

I have seen quite a lot of this coming from China. It is a problem the industry as a whole is facing