vllm-prefix-cache-bench

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

• Third-party content exposure detected (high risk: 0.80). The skill's SKILL.md Option 2 explicitly instructs downloading and ingesting the user-generated ShareGPT dataset from a public Hugging Face URL (https://huggingface.co/datasets/.../ShareGPT_V3_unfiltered_cleaned_split.json), which the benchmarking workflow reads and replays as prompts that can materially influence model outputs and behavior, enabling indirect prompt injection.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill instructs fetching and using the ShareGPT dataset at https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json (which will directly supply prompts used at runtime) and also documents cloning the vLLM repo from https://github.com/vllm-project/vllm (which may be executed via the provided benchmark scripts), so remote content can directly control prompts or execute code.