ai-improving-accuracy

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

• Third-party content exposure detected (high risk: 0.80). The skill explicitly ingests public third‑party data (e.g., "From HuggingFace" using datasets.load_dataset in Step 2 and reference.md) as the dev/train set that the agent reads and uses to drive evaluation and optimization, so untrusted dataset content could influence prompting/behavior.