evalscope-docs

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

• Third-party content exposure detected (high risk: 0.85). Outsider-authored free text is ingested at runtime via benchmark/dataset sample fields (e.g., questions/documents) that are loaded from external datasets (often auto-downloaded) and then inserted into the model prompt/messages context for inference/judging.

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

• Potentially malicious external URL detected (high risk: 0.80). Several benchmark docs (e.g., AIR-Bench and AA-LCR) state EvalScope will lazily download datasets from ModelScope at runtime (e.g., https://modelscope.cn/datasets/evalscope/AIR-Bench-Dataset and https://modelscope.cn/datasets/evalscope/AA-LCR/summary), and those downloaded dataset contents directly control the prompts/samples used for evaluation.