tao-train-optical-inspection
Pass
Audited by Gen Agent Trust Hub on Sep 22, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill generates Docker commands to run the TAO toolkit for machine learning tasks like training, inference, and engine generation.
- [EXTERNAL_DOWNLOADS]: The skill references official NVIDIA Docker images from the trusted container registry
nvcr.io. - [INDIRECT_PROMPT_INJECTION]: The skill ingests external data sources (CSV files and image datasets) for training and evaluation. While this is an attack surface, the risk is mitigated as these operations occur within a containerized machine learning environment.
- Ingestion points: CSV dataset paths and image directories defined in
SKILL.mdandreferences/skill_info.yaml. - Boundary markers: None explicitly defined for the data files themselves.
- Capability inventory: Subprocess execution of
optical_inspectioncommands via Docker. - Sanitization: Relies on the underlying TAO Toolkit's data parsing and validation.
- [SAFE]: The skill includes explicit security guidance, specifically warning users not to set
TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1when using untrusted checkpoints to prevent potential code execution during model loading.
Audit Metadata