skills/nvidia/skills/verl-e2e-testing/Gen Agent Trust Hub

verl-e2e-testing

Pass

Audited by Gen Agent Trust Hub on May 20, 2026

Risk Level: SAFE
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill fetches the verl repository from its official GitHub organization (verl-project) and installs the nvidia-modelopt package from the vendor's package index (pypi.nvidia.com). These sources are well-known and affiliated with the vendor.
  • [COMMAND_EXECUTION]: The skill uses shell commands, Slurm tasks (srun), and Python blocks to manage environment variables (PYTHONPATH), check repository status, and execute orchestration scripts. These actions are standard for the described E2E validation task.
  • [REMOTE_CODE_EXECUTION]: The instructions involve executing a shell script (run_ppo_trainer_megatron.sh) contained within the cloned verl repository to launch the PPO training loop. This is a core component of the legitimate testing workflow.
Audit Metadata
Risk Level
SAFE
Analyzed
May 20, 2026, 05:58 AM
Security Audit — agent-trust-hub — verl-e2e-testing