Illicit model distillation · GTG-16012 · GTG-16003
Allegations of Third-Party Resale and Training Data Ecosystem: Cases GTG-16012/16003
This page is a Chinese-to-English translation of content from pages 152–153 of Anthropic's September 2026 report. All statements regarding specific companies are Anthropic's unilateral allegations; this site has not independently verified them and they do not constitute a finding of illegality.
According to an Anthropic report, the proliferation of proxy services has created a secondary market through which labs can buy or obtain records of users' exchanges with Claude. SenseTime's distillation pipeline included user-Claude exchange records purchased from third-party data vendors. MiniMax built a proxy network service through shell companies that only provided access to Anthropic and OpenAI models, not any Chinese models, suggesting its purpose was to collect users' exchanges with US frontier models to train its own model. All of the above are allegations by the publisher and have not been independently verified by this site.
What happened
The proliferation of proxy services used to evade Anthropic's access restrictions has created a secondary market through which labs can buy or obtain records of users' exchanges with Claude. Some proxy networks both provide Claude access to users in unsupported regions and save exchange records to sell them to other labs.
For example, SenseTime's distillation pipeline included user-Claude exchange records purchased from third-party data vendors. These exchange records were collected from users who accessed Claude through intermediaries such as third-party apps or routing services, which recorded the exchanges and sold them. SenseTime also used Claude to write its distillation pipeline, as well as to launch and monitor training runs.
MiniMax built its own proxy network service through a shell company. This shell company had no obvious connection to MiniMax and did not disclose its relationship with the parent company. The shell proxy network service only provided access to models developed by Anthropic and OpenAI. The service did not provide access to any Chinese models, including MiniMax's own model. This evidence suggests that MiniMax established the proxy network service to collect users' exchanges with US frontier models to train its own model.
What AI did in this case
SenseTime used Claude to write its distillation pipeline, as well as to launch and monitor training runs.
MiniMax built a proxy network through shell companies to collect users' exchange records with US frontier models.
What the report observed
The report confirms that proxy services have created a secondary market through which labs can buy users' exchange records with Claude.
The report confirms SenseTime's distillation pipeline included user-Claude exchange records purchased from third-party data vendors.
The report confirms MiniMax built a proxy network service through shell companies that only provided access to Anthropic and OpenAI models.
Confirmed & unknown
Confirmed
- The report alleges the proliferation of proxy services has created a secondary market through which labs can buy users' exchange records with Claude
- The report alleges SenseTime's distillation pipeline included user-Claude exchange records purchased from third-party data vendors
- The report alleges MiniMax built a proxy network service through shell companies that only provided access to Anthropic and OpenAI models
Unknown
- The report does not state the specific sources and scale of the exchange records purchased by SenseTime
- The report does not state the number of users and volume of exchange records on MiniMax's proxy network service
- The report does not include these companies' responses to the allegations
Platform response
Anthropic states it is strengthening monitoring of proxy services and third-party data vendors.
Limits of response：Banning accounts cannot stop third-party data vendors from continuing to sell users' exchange records.
Takeaways
- Proxy services not only provide model access but may also collect and sell users' exchange records, forming a secondary market for training data.
- Building proxy networks through shell companies can hide the real operator and increase regulatory difficulty.
- When users access models through third-party services, their exchange records may be collected and sold to other labs.