compare-trials

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

• Third-party content exposure detected (high risk: 0.70). The skill’s runtime path ingests the contents of the provided JSONL files (e.g., agent-eval-harness eval ... > graded.jsonl), and those JSONL rows can include outsider-authored free text in fields like input/output that the script parses as readable strings via loadJsonl() → Bun.file(path).text() → JSON.parse(line).