LiteLLM is a proxy server and gateway for accessing large language models through a single controlled entry point. In practice, it can route requests to different model providers, manage API keys, enforce authentication, and expose higher-level features such as tool access for model-driven workflows. By sitting in front of the models, LiteLLM becomes part of the trust boundary for an AI deployment.
That matters because security controls at the gateway determine who can use the models and what they can do. If authentication or authorization fails, an attacker may be able to reach protected model endpoints, misuse billing or API credentials, or invoke connected tools and integrations such as MCP services. Defenders should treat LiteLLM like critical infrastructure: keep it patched, restrict network exposure, review fallback and passthrough authentication logic, and monitor logs for unusual token use or tool requests. In attacks, the gateway is often the easiest place to bypass controls; in defenses, it is where access should be tightly enforced.