external-gitcode-ascend-verl-feature-deploy

Pass

Audited by Gen Agent Trust Hub on May 19, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSCREDENTIALS_UNSAFE
Full Analysis
  • [COMMAND_EXECUTION]: The skill utilizes docker run with the --privileged flag and mounts several sensitive host paths, including /dev/davinci* (NPU devices), /usr/local/Ascend (driver/toolkit paths), and /home. These high-privilege settings are a functional requirement for enabling NPU acceleration and distributed training within containers.
  • [COMMAND_EXECUTION]: Multiple files (assets/start_template.sh, references/ops-commands.md, references/troubleshooting.md) contain commands to terminate processes (pkill -9 python, ray stop --force) and remove temporary files (rm -rf /tmp/ray). In the context of ML operations, these are standard cleanup and recovery procedures used to reset the environment for new training runs.
  • [EXTERNAL_DOWNLOADS]: The skill pulls Docker images from quay.io/ascend/verl and installs the swanlab Python package. These downloads originate from well-known registries and are directly related to the skill's primary purpose of model training and monitoring.
  • [CREDENTIALS_UNSAFE]: The skill prompts the user for a SwanLab API Key. This credential is used for a legitimate login operation via the official SwanLab CLI (swanlab login) and is stored in the service's default configuration directory (~/.swanlab/). The handling of the key follows standard practices for this integration.
Audit Metadata
Risk Level
SAFE
Analyzed
May 19, 2026, 06:27 AM
Security Audit — agent-trust-hub — external-gitcode-ascend-verl-feature-deploy