A vast network of intermediary servers is enabling users in China to access frontier AI models in the US without revealing their identities and location, likely to circumvent potential access restrictions and monitoring.
The volume of activity flowing through the network suggests systematic attempts to use outputs from frontier AI models to create less capable versions at substantially lower cost than developing them independently, Team Cymru said in a report this week. The findings come just days after the US government accused Chinese AI companies of attempting to clone US AI capabilities via systematic distillation campaigns.
A Sprawling Network of 80K LLMs
Researchers from the company recently identified more than 80,000 LLM relay servers that are acting as intermediaries between users and AI providers. The relay servers allow operators to pool credentials for multiple AI accounts and route user requests through the servers to frontier services, including Anthropic, OpenAI, Google, and xAI.
The proxies can undermine the controls that frontier AI providers rely on to detect and restrict misuse, Team Cymru researcher Scott Fisher wrote in the analysis. By separating the apparent requester from the actual user, relay servers can help obscure who is accessing models and enable users to bypass potential geographic restrictions. They also allow users to share or resell credentials, and to evade a provider’s terms of service.
“A transfer station breaks the assumption every frontier-model control depends on: that the account making a request belongs to the party consuming the answer,” Fisher wrote. “Account attribution, usage metering, rate limits, abuse detection, regional availability, and the terms that govern all of them are enforced against whoever holds the credentials.” Relay servers break this assumption and enable fraud on a massive scale, he said.
High-Volume Chinese Traffic to OpenAI, Others
Team Cymru initially identified a total of 10,867 transfer stations across 457 autonomous systems, or ASNs, but updated the number to more than 80,000 relays after further digging. The company examined a cluster of relay servers hosted by several US virtual private server providers and found more than 4,000 IP addresses in China and Hong Kong connecting to them. Over an eight-day period, those addresses sent about 14TB of data to the relay stations and received more than 7TB in return. Some of the same relays connected to both Chinese AI providers and Western companies, including Anthropic, OpenAI, Google, and xAI.
According to Team Cymru, the traffic it observed going directly to an Anthropic API was especially noteworthy for its sheer volume. Over the eight day period, users uploaded some 81GB of data to Anthropic via just 17 relay servers and downloaded about 1.4GB in responses. Team Cymru interpreted the 58:1 upload-to-download ratio as potentially consistent with large-scale model distillation attempts where attackers use a frontier model’s outputs as training data to develop a cheaper, less capable model that mimics its behavior. “Chinese users and organizations are, at minimum, violating the access blocks to the service, and can also easily steal output to clone their own versions,” Fisher said.
Team Cymru identified two open source software packages, Claude Relay Service (CRS 1.x) and its successor, sub2api; both were published by developer Wei-Shaw on GitHub in the Simplified Mandarin that many of the relay servers are using as proxies for accessing frontier models.
The newer sub2api software supports user management, per-user billing, subscription-to-API conversion, and prompt auditing. Team Cymru’s analysis found that users had forked the software more than 8,000 times on GitHub, while its Telegram channel had nearly 7,000 subscribers. The figures point to widespread interest in the software, though they don’t establish how many people are actively using it for malicious purposes.
From a functionality standpoint, Claude Relay Service and sub2api let operators combine multiple AI accounts or API credentials into a shared gateway. Operators pool API keys or logged-in subscription accounts and issue their own credentials to customers, routing requests through the relay so AI providers see the relay’s IP address and upstream credentials rather than the end user’s IP address or direct identity.
Team Cymru found on sub2api’s GitHub page a listing of 26 commercial sponsors, of which 15 sold access to models such as Claude and OpenAI, seven provided IP addresses, two supplied accounts that people could use to access AI services, and one each provided AI API traffic-optimized infrastructure and services for AI image and video generation.

Comments are closed