serving-llms-vllm

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

• Third-party content exposure detected (high risk: 0.90). The skill's workflows explicitly instruct fetching and using public models from HuggingFace/TheBloke (e.g., "Step 2: Find pre-quantized model" in references/quantization.md and example launches like TheBloke/Llama-2-70B-AWQ) and even recommend using --trust-remote-code in references/server-deployment.md, which clearly exposes the agent to untrusted third-party model code and user-provided content that can influence behavior.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill instructs vllm to load remote models (e.g., TheBloke/Llama-2-70B-AWQ / meta-llama/Llama-3-8B-Instruct) at runtime and explicitly shows use of the --trust-remote-code flag, which causes code from the model repository (e.g., https://huggingface.co/TheBloke/Llama-2-70B-AWQ) to be fetched and executed, so remote content can run code and directly affect agent behavior.