Researcher Chaofan Shou, co-founder of the blockchain security firm Fuzzland, reported acquiring approximately 6 terabytes of model call data from a Chinese large language model relay, also described as a router or proxy. In these logs, he discovered sensitive credentials.
In messages and secondary reports circulated on September 11th, including discussions on X trends, Shou stated that the material contained SSH keys, VPN configurations, Alibaba Cloud keys, and GitLab tokens. He claimed this was sufficient to gain access to servers or internal systems associated with about 19 Chinese companies and several state-affiliated organizations. Examples mentioned in these reports included Huawei, Xiaomi, NIO, and MiniMax.
Pandaily has not independently verified this dataset, the validity of the keys, or any successful intrusion; therefore, this article treats this case as a reported research finding rather than a confirmed multi-corporate breach.
Technically, this is a known intermediary risk in the AI application supply chain. LLM routers sit between users and upstream models, often viewing the full text of prompts and responses. When developers place infrastructural secrets in agent contexts—such as keys, tokens, VPN profiles, or repository credentials—the relay that stores or resells the logs can turn a convenient layer into a credential harvesting point for enterprises and internet companies relying on third-party routing.
Previously, Shou co-authored research on malicious attacks via intermediaries across hundreds of relays, reporting instances of injected tool calls, opportunistic use of embedded cloud service test keys, and wallet redirection in controlled tests. These earlier findings support the structural thesis, even if the new 6 TB acquisition remains a researcher's claim rather than a disclosure package from the named firms.
Public information currently lacks details on the relay operator, how long the records were stored, or whether the credentials were active at the time of verification. The reports analyzed for this article did not contain broad confirmations, forced rotations, or official incident notifications from the mentioned enterprises.
The practical takeaway for providers and users of Chinese LLM proxies is to implement defensive and operational measures: treat relays as high-trust nodes transmitting data in plain text; exclude secrets from agent prompts; rotate any keys that may have passed through third-party routers; and prefer audited gateways with strict logging and storage controls.