Did you think you were communicating with DeepSeek or Kimi? Surprise! It was often Claude from Anthropic, behind the scenes. This revelation uncovers a dangerous game among AI giants.
Claude from Anthropic: the discreet star hijacked by its Chinese competitors
Claude, Anthropic’s AI model, has become the star of a strange industrial soap opera. Since December, seven Chinese labs, including Alibaba, DeepSeek, Moonshot AI, and Xiaomi, have reportedly diverted nearly 200 million exchanges, according to an explosive report from Anthropic published in early September. This industrial siphoning operation does not resemble a simple classic imitation.
The technique used is called “distillation”: a smaller model learns to respond based on the answers of a larger model, the “teacher”. It’s a standard method, but here, the “teacher” didn’t ask for it! The Chinese labs have therefore made Claude respond instead of their own services, then displaying the answers as their own creations.
For example, Moonshot AI transferred 300,000 client requests to Claude through over 5,000 fraudulent accounts based in Singapore and Japan. DeepSeek, on the other hand, developed a more sophisticated mechanism. It captured requests from developers using tools like Claude Code and secretly redirected them to Claude Opus.
The massive data theft at the heart of an industrial chain
Alibaba doesn’t do things by halves, with a peak of 151 million intercepted exchanges in just three months. This represents up to 3 million interactions per day, orchestrated through more than 3,500 fake accounts, using stolen credit cards and phantom identities.
Zhipu and Xiaomi complete this chaotic picture by re-injecting user sessions into Claude, while SenseTime procures transcripts through resellers. MiniMax, for its part, operates through a shell company selling only American models, with no Chinese models… a subtle strategy.
This true cascading traffic shows the scale of the logistics chain of AI theft. Intermediaries seize conversations, extract reasoning, and then feed their own systems.
Impact and stakes beyond rivalries between labs
This conflict is no longer limited to a simple game of dupes among competitors. Since 2025, OpenAI and Google have also pointed out this phenomenon of illicit distillation. What changes this time are the proofs of an organized system encompassing API theft and credit card fraud. Anthropic thus speaks of “practices probably incompatible with GDPR,” as sensitive data circulates via European platforms without control.
The risks go beyond mere “copy-paste”. The report reveals sensitive data, including email addresses, still valid access tokens, and even investment forecasts from a pharmaceutical lab. More shocking: a user, presumably linked to the Chinese army, allegedly sent surveillance videos to analyze, thinking he was interacting with Kimi, while he was actually face-to-face with Claude.
For advocates of transparency, this case revives the urgency to strengthen digital sovereignty and protect the integrity of AI models. Trust no longer arises from a simple declaration, but from the transparency of mechanisms and open control of data.
Innovative protection methods by Anthropic against data theft
In the face of these attacks, Anthropic has revisited its approach and implemented robust measures. The Claude model now incorporates a reasoning summary system to reduce the value of stolen data. Furthermore, invisible marking solutions are under study to trace and authenticate generated texts, thus limiting abusive reuse.
This digital trust battle is not isolated. Many experts are interested in this invisible marking — a promising avenue for authenticating and securing generated content.
You can find detailed information about these innovations on this dedicated page on Anthropic’s traceability solutions or on emerging digital marking methods at industry leaders.
Key points revealed by Anthropic’s report on DeepSeek and Kimi
- Over 200 million diverted exchanges, targeting seven Chinese labs including Alibaba and Xiaomi.
- Massive use of fraudulent accounts with identity theft and credit card fraud.
- Secret redirection of user requests to Claude, without users knowing.
- Theft of sensitive information including military and industrial data.
- Illegal logistics chain of intermediaries reselling access to Claude on a large scale.
To dive deeper into this complex case
The complete report from Anthropic is accessible, but for a thorough synthesis, you can read a clear and critical file on the wild distillation of Chinese AIs. A fascinating and worrying context that the free digital world cannot ignore.
Another major resource, quantumzeitgeist.com details precisely the hijacking techniques used in this case.
This video offers a complete insight into the scope and methods of the hijacking.
Watch this analysis to understand AI distillation from a technical angle and its implications for digital sovereignty.
What is ‘distillation’ in AI?
Distillation is a training method where a smaller model learns to reproduce the responses of a larger model, allowing to lighten the computational load.
Why does this hijacking pose a problem?
Because it exposes unprotected sensitive data and steals the work of an AI model without permission, widening the gap of unfair practices and raising confidentiality issues.
How does Anthropic try to protect itself?
Anthropic has implemented reinforced filters, a summary of Claude’s reasoning, and is developing an invisible marking system to authenticate its content.
Are users aware?
Often, users do not know they are ultimately interacting with Claude while thinking they are using DeepSeek or Kimi.
Could this scandal have a legal impact?
Yes, because it involves intellectual property, personal data, and questions the compliance with GDPR, particularly concerning data circulating via European platforms.
Source: www.01net.com