lean4-prove

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

• Third-party content exposure detected (high risk: 0.90). The SKILL.md and task docs explicitly describe ingesting and retrieving 94k+ exemplars (via ./ingest.sh and ArangoDB) and benchmark.py/converge.py load public HuggingFace datasets (e.g., "cat-searcher/minif2f-lean4", "deepseek-ai/DeepSeek-ProverBench") so untrusted, public third‑party content is fetched and directly used to constrain LLM generation and compilation decisions.