tao-train-metric-learning-recognition

Pass

Audited by Gen Agent Trust Hub on Sep 17, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill references official NVIDIA TAO Toolkit container images from the nvcr.io registry for training and deployment actions.
  • [COMMAND_EXECUTION]: Provides instructions for docker run commands to initiate model training and deployment processes. These commands are fundamental to the skill's documented use case.
  • [DYNAMIC_EXECUTION]: The skill documents the environment variable TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1, which enables full object loading in PyTorch. The documentation includes an explicit warning to use this setting only with trusted checkpoints to mitigate potential risks associated with insecure deserialization.
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests dataset paths and configuration parameters. While this presents a potential attack surface, the operations are performed within the context of specific, well-defined machine learning workflows.
  • Ingestion points: Dataset paths such as dataset.train_dataset and dataset.val_dataset in SKILL.md and references/skill_info.yaml.
  • Boundary markers: None identified.
  • Capability inventory: Executes shell commands and interacts with local filesystems via the Bash and Read tools.
  • Sanitization: No explicit sanitization or validation of the external dataset contents is performed within the skill instructions.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 17, 2026, 02:56 AM
Security Audit — agent-trust-hub — tao-train-metric-learning-recognition