devtu-benchmark-harness

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

• Third-party content exposure detected (high risk: 0.85). At runtime, scripts/run_eval.py reads BixBench capsule files (e.g., capsule_path contents) and injects their readable text into the LLM prompt via prepare_prompt() and precompute_for_capsule() (e.g., CSV/metadata-derived outputs and “Data files are located at…” + file listings), so the LLM context can include outsider-authored free text from the capsule dataset.