nemo-gym-pivot-datasets

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

• Third-party content exposure detected (high risk: 0.75). The required workflow for this skill (“create local single-step pivot datasets from successful trajectories” and “convert each accepted decision point into one pivot JSONL row”) ingests outsider-authored free text from the provided source rollout/trajectory artifacts (e.g., JSONL/chat messages containing system/user/assistant/tool text) and materializes it into LLM-readable fields like responses_create_params.input and expected_action; the supplied SKILL.md itself does not perform the ingestion, but the runtime converter/validator workflow explicitly reads those artifacts and writes them into the dataset structure that is then consumed by an LLM.