dataset-evaluation

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

• Third-party content exposure detected (high risk: 0.70). Outsider free text can be ingested when the dataset is provided as an S3 URI: scripts/format_detector.py reads the first 1MB of the file via _sample_s3_file() and parses each JSONL line (json.loads(line)), so any outsider-authored dataset content becomes LLM-readable text in the agent’s context through the tool’s validation output/errors.