serving-llms-vllm

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.85). Outsider free text can enter the LLM context via the runtime OpenAI-compatible request body (e.g., messages[].content / prompt) sent by external callers to the vLLM server (vllm serve ...), which the server forwards into generation.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.80). The skill runs vllm serve with remote model names (e.g., meta-llama/Llama-3-8B-Instruct and TheBloke/Llama-2-70B-AWQ) which causes vLLM to fetch artifacts from HuggingFace (e.g., https://huggingface.co/meta-llama/Llama-3-8B-Instruct) and the documentation explicitly references using --trust-remote-code, which can execute remote repository code at runtime.