Test whether content planted in a RAG corpus (or otherwise retrieved) can steer the model's answers or trigger actions — i.e. indirect prompt injection and data poisoning via the retrieval path. Use on an authorized RAG app to validate ingestion/retrieval trust boundaries.
