nemo-gym-pivot-datasets
Pass
Audited by Gen Agent Trust Hub on Jul 22, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The validation scripts (scripts/validate_pivot_dataset.py and scripts/reference/tool_messages_to_pivot_dataset_reference.py) dynamically modify the Python path (sys.path) using a user-provided argument (--gym-repo). This allows the script to import and execute modules from arbitrary local directories.\n- [PROMPT_INJECTION]: The skill processes untrusted dataset files (JSONL) which contain assistant and user messages. An agent using this skill is instructed to 'inspect' these rows, creating a vector for indirect prompt injection where malicious instructions embedded in the data could influence the agent's behavior.\n
- Ingestion points: JSONL file reading in conversion and validation scripts.\n
- Boundary markers: Absent; the agent is not warned to ignore instructions within the data rows.\n
- Capability inventory: Local file access, directory modification, and dynamic module loading.\n
- Sanitization: The scripts validate JSON structure but do not sanitize the semantic content of the messages.
Audit Metadata