rerun-lerobot
Pass
Audited by Gen Agent Trust Hub on Oct 14, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches dataset content and metadata from the Hugging Face Hub (huggingface.co) using the official huggingface_hub library. This interaction targets a well-known service for robotics data.
- [COMMAND_EXECUTION]: Instructs the agent to run bundled scripts (scripts/find_lerobot_dataset.py, scripts/fetch_lerobot_episodes.py) via uv run. These scripts perform localized file management and network tasks appropriate for the skill's purpose.
- [INDIRECT_PROMPT_INJECTION]: The skill processes metadata from external Hugging Face repositories, creating a surface where malicious dataset metadata could potentially influence agent operations.
- Ingestion points: The Python scripts ingest meta/info.json and meta/episodes.jsonl files from remote repositories using the huggingface_hub API.
- Boundary markers: No delimiters or explicit instructions are used to separate untrusted metadata from the agent's instructions.
- Capability inventory: The skill possesses capabilities to write to the local filesystem (Path.write_text), delete directories (shutil.rmtree), and download remote assets (snapshot_download).
- Sanitization: Destination paths are sanitized by replacing directory separators in repository IDs, and metadata is handled via structured JSON parsing without dynamic evaluation.
Audit Metadata