Artificial intelligence routers that route consumer requests between different models can pose serious security risks. A researcher discovered terabytes of data containing potentially dangerous secrets in a leak.
According to researcher Chaofan Shaw, approximately 6 terabytes of dumps collected from Chinese routers contain session data related to Anthropic’s Claude. These logs store requests, responses, tool calls, and metadata, and developers often use them to transmit sensitive information. This may include SSH keys, VPN settings, cloud API keys, and GitLab tokens. Therefore, a major breach could open access to multiple systems simultaneously.
Xiao claimed that the data discovered could impact government agencies in China and the CIS, as well as major companies such as Xiaomi, Huawei, NIO and Minimax. He believes the size of the leak shows real production logs rather than a prepared set of information. Researchers have previously studied man-in-the-middle attacks in artificial intelligence routers that can block secrets in bulk.
The story got more context after Anthropic reported that the company accused Chinese artificial intelligence labs Moonshot and DeepSeek of redirecting requests to Claude. Another researcher expressed skepticism about this version, arguing that on-the-fly inference from the service is not well suited to the situation. The necessary data may have arrived through the router itself.
