analytical-method-validation

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

• Third-party content exposure detected (high risk: 0.75). The required runtime workflow ingests outsider-authored free text in the form of user-provided CSV/TSV/JSON files (e.g., --input, --run, --isr) via read_input() as UTF-8 strings and parse_rows() into LLM-visible stdout/stderr-friendly prose (“finding: …”), so attacker-controlled text can indirectly reach the agent’s context.