Measuring Malicious Intermediary Attacks on the LLM Supply Chain

Large language models (LLMs) increasingly depend on third-party API routers that act as application-layer proxies. These routers have full plaintext access to JSON payloads but lack cryptographic integrity enforcement. This study systematically examines this attack surface, formalizing a threat model for malicious LLM API routers and defining two main attack classes: payload injection and secret exfiltration. The findings reveal that several routers actively inject malicious code and can exploit benign configurations, highlighting significant security risks for LLM deployments.