verl-rl-training

Warn

Audited by Gen Agent Trust Hub on Mar 28, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The skill provides instructions for executing complex shell commands to initiate distributed training and manage Ray clusters. It also supports the dynamic loading of arbitrary code via the custom_reward_function.path configuration, which imports logic from local Python files for reward computation.
  • [PROMPT_INJECTION]: The training workflows ingest large datasets from external Parquet files, creating a surface for indirect prompt injection where malicious data could influence the training process or reward estimation.
  • Ingestion points: Training and validation files specified in the data.train_files and data.val_files configuration parameters.
  • Boundary markers: No explicit boundary markers or isolation mechanisms are mentioned for separating instructions from untrusted dataset content.
  • Capability inventory: The skill executes shell commands, writes to the filesystem, and dynamically loads Python modules at runtime.
  • Sanitization: No explicit sanitization or verification of dataset content is demonstrated in the provided examples.
  • [EXTERNAL_DOWNLOADS]: The instructions involve cloning the verl source code from the official GitHub repository of Volcengine and pulling Docker images from the verlai organization on Docker Hub.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 28, 2026, 06:07 PM