UC Study Reveals Vulnerabilities in LLM Routers

University of California researchers revealed significant security vulnerabilities in third-party large language model routers on Thursday, warning that these intermediaries can inject malicious code and steal cryptocurrency credentials, potentially exposing developers to financial losses.

The study, which examined malicious intermediary attacks in the LLM supply chain, tested 28 paid routers and 400 free ones sourced from public developer communities. Researchers found that nine of the services actively inserted harmful code into AI workflows, while 17 attempted to access the team's Amazon Web Services credentials. In one controlled experiment, a router successfully extracted a private key and drained Ether from a decoy wallet, resulting in a loss of less than $50. Overall, 26 routers demonstrated behaviors that enabled credential theft or code injection, according to the findings.

These routers, which developers use to route API requests to providers like OpenAI, Anthropic, and Google, terminate Transport Layer Security connections. This design grants them full access to unencrypted data, including sensitive information such as private keys and seed phrases. Developers relying...