tao-run-deft-object-detection
Pass
Audited by Gen Agent Trust Hub on Sep 26, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: Comprehensive static analysis of the skill's instructions, bundled Python scripts, and reference markdown files revealed no security risks, malicious patterns, or unauthorized data access behaviors.
- [COMMAND_EXECUTION]: The skill uses the
subprocessmodule to executedocker runcommands and thengcCLI. These operations are required for the skill's core functionality—executing model training, inference, and analytics tasks within isolated containers, and downloading official model weights. Commands are constructed using verified paths and state-backed parameters. - [EXTERNAL_DOWNLOADS]: The skill facilitates the download of official pre-trained model checkpoints from the NVIDIA NGC (GPU Cloud) registry and references model encoder snapshots from HuggingFace. These are well-known, trusted industry services, and the downloads are limited to the model artifacts specified in the workflow.
- [DYNAMIC_EXECUTION]: To manage the complex multi-stage pipeline, the skill programmatically generates and updates YAML specification files for various TAO toolkit tasks. This is implemented using structured YAML libraries (PyYAML) and template-based overrides to ensure configuration consistency and prevent command injection during spec generation.
Audit Metadata