tao-train-visual-changenet

Pass

Audited by Gen Agent Trust Hub on Jun 23, 2026

Risk Level: SAFE
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill instructs users to download pretrained weights from HuggingFace (nvidia/C-RADIOv2-B) and datasets from S3 buckets. These are standard operations for machine learning workflows and target well-known, legitimate services.
  • [COMMAND_EXECUTION]: The skill provides templates for executing docker run commands and scripts to resolve container versions. These are restricted to the intended purpose of running the TAO Toolkit container.
  • [COMMAND_EXECUTION]: Instructions are included for compiling custom CUDA operators (MultiScaleDeformableAttention) within the container environment, which is a required step for specific segmentation model backbones.
  • [SAFE]: No instances of prompt injection, data exfiltration, or obfuscation were found. Authentication for external services is handled through standard environment variables and local configuration files without hardcoding secrets.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 23, 2026, 12:09 PM