serving-llms-on-epyc

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

• Third-party content exposure detected (low risk: 0.10). Outsider-authored free text only reaches runtime via the user-supplied --model-id parameter, but the required workflow does not ingest arbitrary external text feeds/queues; the scripts read first-party/trusted configs/metadata and then make network GETs specifically tied to that model ID (e.g., HuggingFace config.json and repo registry.py), not from an unsolicited outsider posting mechanism.

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

• Potentially malicious external URL detected (high risk: 0.80). The runtime scripts (e.g., scripts/check_model.py and scripts/estimate_memory.py) perform HTTP GETs to external resources—https://raw.githubusercontent.com/vllm-project/vllm and https://huggingface.co—which are fetched at runtime and directly influence the agent's decisions about model support and memory fit.