tao-run-deft-aoi
Pass
Audited by Gen Agent Trust Hub on Jun 30, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The skill makes extensive use of
docker runto execute various components of the NVIDIA TAO Toolkit, including training, inference, and data services. This is a core requirement for the Visual ChangeNet workflow. - [EXTERNAL_DOWNLOADS]: Downloads model weights, pretrained backbones, and reference datasets from trusted sources including HuggingFace (e.g.,
nvidia/C-RADIOv2-B) and NVIDIA's NGC registry (nvcr.io). These downloads are necessary for the skill's functionality and target well-known, official repositories. - [COMMAND_EXECUTION]: Uses the
chmod 777command on specific workspace subdirectories (e.g.,augmentation/anomalygen/) to ensure the Docker container, which runs with a non-root user ID, has sufficient permissions to write output data and checkpoints back to the host system. - [COMMAND_EXECUTION]: Includes a script (
scripts/align_token_usage.py) that accesses local agent transcripts stored in~/.claude/projects/. This access is used to perform post-run token usage accounting, which is then reflected in the generated HTML report to provide the user with cost transparency. - [COMMAND_EXECUTION]: Orchestrates multiple Python scripts to perform data validation, CSV manipulation, and state management. These scripts use standard libraries (e.g.,
pathlib,json,csv) to ensure the integrity of the training data and execution state.
Audit Metadata