tao-validate-dataset-format

Pass

Audited by Gen Agent Trust Hub on Sep 17, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external datasets, which introduces a surface for indirect prompt injection if the datasets contain malicious metadata or instructions.
  • Ingestion points: The skill ingests untrusted data from the directory or file specified via the --path argument during the execution of tao-daft validate.
  • Boundary markers: The instructions do not provide explicit boundary markers or instructions to ignore embedded content within the processed datasets.
  • Capability inventory: The skill leverages the Bash tool to execute the validation command and requires access to the local file system to read dataset files.
  • Sanitization: There is no mention of input sanitization or content filtering for the data being validated within the skill's instructions.
  • [EXTERNAL_DOWNLOADS]: The skill documentation prescribes the installation of a Python dependency.
  • Evidence: Instructions in SKILL.md and skill-card.md recommend running pip install nvidia-tao-daft to install the necessary validation tools.
  • Note: The nvidia-tao-daft package is a resource provided by the vendor (NVIDIA) for use with their TAO Toolkit.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 17, 2026, 02:55 AM
Security Audit — agent-trust-hub — tao-validate-dataset-format