evaluating-llms-harness

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

• Third-party content exposure detected (high risk: 0.80). This skill explicitly ingests public third‑party content — e.g., HuggingFace datasets and user-provided JSONL/CSV files via references/custom-tasks.md and arbitrary API responses via references/api-evaluation.md's TemplateAPI/base_url — and the harness reads and interprets those datasets and API outputs as part of its evaluation workflow, enabling indirect prompt injection from untrusted sources.