evaluate-ghostwriter

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

• Third-party content exposure detected (high risk: 0.85). The runtime path scripts/run-eval.ts reads outsider-authored case-data JSONL (from evals/cases.jsonl/--cases), converts it into a “Case data” prompt string via buildCommonPrompt()/buildCandidatePrompt(), and feeds that free text into the model invocation; this is an outsider free-text source (cases are not created by the user per SKILL.md).