6 TB of data leaked from Chinese LLM router, exposing secret keys

AI security researchers have encountered another serious issue. According to ixbt.com, a massive database of nearly 6 TB of data from Chinese LLM router sessions was found exposed, containing secret access keys belonging to major tech giants and government institutions. This incident highlights the potential risks associated with AI intermediaries. reports Ixbt.com.

In the modern digital ecosystem, LLM routers distribute requests from users or AI agents across various language models. They receive the request, select the appropriate neural network, and transmit the data. However, during this process, the service may store detailed logs, which include not only text queries and responses but also various technical metadata and external tool calls.

Leak of sensitive data via intermediaries

According to an analysis by researcher Chaofan Shou, the main risk arises when developers or autonomous AI agents transmit confidential data to models to perform tasks. It has become common for requests to contain SSH keys, VPN configurations, cloud API keys, GitLab tokens, and other secrets providing access to real information systems.

If the intermediary distributing the data does not sufficiently protect its traffic, its logs essentially become a repository of ready-to-use secrets. This is exactly what happened with one of the LLM routers in China, where requests sent to Anthropic's Claude AI were also found stored.

Major companies and government agencies at risk

According to the researcher, the 6 TB array contained active secret keys allowing access to the systems of Chinese government organizations and large tech companies. Brands such as Xiaomi, Huawei, NIO, and MiniMax were mentioned among those at risk.

Experts note that the presence of functional authentication keys makes such leaks extremely dangerous. An attacker does not need to search for vulnerabilities in a company's infrastructure, as all the data required for system access is already exposed in the intermediary's logs. This incident underscores the need to strengthen security protocols when using AI tools in corporate environments.