skills/robium-ai/robium/lerobot/Gen Agent Trust Hub

lerobot

Pass

Audited by Gen Agent Trust Hub on Sep 4, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONREMOTE_CODE_EXECUTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes LeRobotDataset artifacts which contain natural language task descriptions and iterates through simulation environment outputs that could influence agent behavior.
  • Ingestion points: Ingests natural language strings from meta/tasks.jsonl within datasets and observation data from simulation environments.
  • Boundary markers: No specific delimiters or "ignore embedded instructions" warnings are documented for the processing of task metadata.
  • Capability inventory: The skill performs subprocess execution (uv run), filesystem writes for training outputs, and network operations to the Hugging Face Hub.
  • Sanitization: The analysis did not identify explicit sanitization or filtering of external task descriptions before they are processed by the training or evaluation pipelines.
  • [REMOTE_CODE_EXECUTION]: The EnvHub feature enables the execution of remote Python code by loading simulation environments directly from the Hugging Face Hub.
  • Evidence: Use of make_env(..., trust_remote_code=True) in references/eval-and-sim.md to load community-contributed or official environments.
  • Source Analysis: The feature is configured to target repositories on Hugging Face, which is a well-known service and trusted organization for AI models and datasets.
  • [DYNAMIC_EXECUTION]: Employs library mechanisms to dynamically load and execute environment logic from external repositories at runtime.
  • Evidence: lerobot.envs.make_env allows specifying a repository ID to instantiate environments not locally installed.
  • [COMMAND_EXECUTION]: The skill makes extensive use of CLI utilities for robot training, evaluation, and hardware management.
  • Evidence: Instructions include the use of lerobot-train, lerobot-eval, lerobot-record, and system-level package installation commands like sudo apt install ffmpeg.
  • [EXTERNAL_DOWNLOADS]: Fetches required datasets, model checkpoints, and configuration files from remote servers.
  • Evidence: Downloads resources from huggingface.co and github.com/huggingface/lerobot.
  • Context: These downloads originate from a well-known technology provider and a recognized open-source organization.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 4, 2026, 04:08 AM
Security Audit — agent-trust-hub — lerobot