tao-train-visual-changenet
Pass
Audited by Gen Agent Trust Hub on Sep 17, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill follows standard NVIDIA TAO Toolkit procedures for model training and deployment. All script behaviors and command patterns are consistent with the model's stated purpose of AOI defect detection.
- [EXTERNAL_DOWNLOADS]: The skill fetches pre-trained model weights from NVIDIA's official Hugging Face repository and configuration data from NVIDIA-owned S3 storage. These interactions target well-known and trusted vendor infrastructure.
- [COMMAND_EXECUTION]: Provides instructions to run TAO Toolkit CLI tools (e.g.,
visual_changenet) within a Docker container. These commands are standard for managing the model lifecycle (train, evaluate, export). - [DYNAMIC_EXECUTION]: Documented behavior includes the compilation of custom CUDA kernels required for the segmentation task backbone. This is a routine requirement for specific high-performance model architectures and is executed within the controlled container environment.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests user-provided CSV files for dataset mapping. Risks of path traversal or malformed data are mitigated by the
validate_vcn_dataset.pypre-flight script, which enforces relative paths and specific schema constraints.
Audit Metadata