tao-run-deft-pas
Pass
Audited by Gen Agent Trust Hub on Oct 1, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill fetches pinned TAO Toolkit container images (7.2.0-pyt and 7.2.0-data-services) from NVIDIA's official registry (nvcr.io) and model weights from Hugging Face.
- [COMMAND_EXECUTION]: The workflow involves executing shell commands (e.g., tar, sha256sum) and TAO CLI tools (clip, embedding, tmm) through multiple Python wrapper scripts and shell adapters. These are used for pipeline orchestration and are tightly controlled by status-tracking and audit mechanisms.
- [DYNAMIC_EXECUTION]: The skill uses
os.execveinscripts/run_deft_cli.pyto replace the current process with a verified TAO CLI entrypoint. It also uses asitecustomize.pypatch that dynamically injects an import hook to allow specific NumPy types during PyTorch checkpoint loading (addressing weights-only loading restrictions in newer Torch versions). - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of PAS dataset archives (
images_raw.tar,meta.tar.gz). While these archives provide the training surface, the skill implements strict structural validation, checksum verification, and iteration-bound artifact tracking to mitigate risk.
Audit Metadata