Chinese companies Moonshot AI and DeepSeek secretly redirected user requests to Claude
Quick Look
- Chinese AI companies Moonshot AI and DeepSeek redirected user queries to Anthropic's Claude model, presenting the results as their own.
- Moonshot forwarded nearly 300,000 requests in ten days containing sensitive data, including videos and credentials.
- Anthropic accuses them of industrial distillation, and the US has brought charges against the Chinese developers.
AI-generated summary
Why It Matters
AI model distillation is a method of training a less capable model with the results of a more powerful one. Although technically legal, its use to access other people's computing resources is considered unethical and may violate the models' terms of use.
Chinese artificial intelligence companies Moonshot AI and DeepSeek secretly redirected their users' queries to Claude, presenting the results as answers from their AI models, according to the Anthropic report.
"We discovered that Moonshot AI, the company behind the Kimi family of models, was quietly routing customer requests to Claude instead of processing them through Kimi. Moonshot then displayed Claude's responses to users. These users believed they were using the Kimi model, but instead received responses from Claude," the report said.
As the company points out, in one case, within ten days, Moonshot forwarded almost 300 thousand customer requests to Anthropic. The Chinese company used a network of 5,380 accounts to do this, most of which were located in Singapore and Japan, according to Anthropic’s research.
Developer Claude also claims that the user requests that Moonshot sent to the AI model contained sensitive information. Anthropic does not know whether Moonshot notified its clients that they were forwarding their requests. This information included recordings from CCTV cameras in the city of Chengdu, as well as the internal code and valid credentials of several large Chinese companies, including well-known technology companies.
In its report, Anthropic also notes that the Chinese company DeepSeek used tactics similar to Moonshot. According to the investigation, DeepSeek redirected requests from users attempting to use one of its AI models through third-party Anthropic-owned programming tools such as Claude Code, Claude Agent SDK and OpenCode.
American law enforcement agencies on Tuesday accused Chinese artificial intelligence developers of industrial distillation of advanced AI models created in the United States.
Distillation is a method of training artificial intelligence models that involves training a less capable model based on the results of a more efficient one. Despite the fact that the method is formally widely used and legal, with its help companies can gain access to the capacity of competing laboratories.
Liu Chang, a representative of the Chinese Embassy in Washington, told RIA Novosti that US accusations against China about the alleged “borrowing” of artificial intelligence technologies are unfounded and politically motivated.
What to Watch
AI outlook — possibilities, not facts
The US may impose restrictions on Chinese companies' access to American AI technologies
Likely · Within months
Anthropic will strengthen measures to protect its AI model Claude from unauthorized use
Very likely · Within weeks
China may tighten controls on data exports and use of foreign AI models by domestic companies
Possible · Within months
Open Questions
- Did Moonshot AI notify its clients that their requests were being redirected?
- Which major Chinese companies were affected by credentials and code leaks?
- What measures does Anthropic plan to take to prevent similar incidents in the future?
- Will the US apply sanctions against Moonshot AI and DeepSeek?
