serving-llms-vllm

CRITICAL E006: Malicious code pattern detected in skill scripts.

• Malicious code pattern detected (high risk: 0.90). The docs repeatedly instruct users to enable --trust-remote-code (and show it in production examples), which permits executing arbitrary code from remote model repositories and therefore introduces a high risk of remote code execution/backdoor abuse; no explicit data exfiltration code was found in the docs.

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

• Third-party content exposure detected (medium risk: 0.30). In the vLLM OpenAI-compatible server workflow, outsider-authored user text is read from incoming API requests (e.g., messages[].content / prompt) by the /v1 completion/chat endpoints that call llm.generate with those strings.

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

• Potentially malicious external URL detected (high risk: 0.80). The skill shows running vllm with remote model identifiers (e.g., meta-llama/Llama-3-8B-Instruct and TheBloke/Llama-2-70B-AWQ) which vLLM will fetch at runtime and the docs explicitly mention using --trust-remote-code to allow executing code from those remote model repositories, creating a high-confidence remote-code execution risk.