distill-test-set

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

• Third-party content exposure detected (high risk: 0.70). When --source is coval:<TEST_SET_ID>, scripts/distill.py calls coval_pull() to fetch every test case via the Coval API and then ingests the outsiders’ input/expected text into the LLM context indirectly via the generated distilled.ndjson (which is later bulk-loaded); this is outsider-authored content from another party’s test set, not user-authored.