tao-train-nvdinov2
Pass
Audited by Gen Agent Trust Hub on Oct 29, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill consists of documentation, JSON schemas, and spec templates for the NVIDIA TAO Toolkit. No evidence of malicious code, prompt injection, or unauthorized data access was found.\n- [INDIRECT_PROMPT_INJECTION]: The skill facilitates the processing of large image datasets for self-supervised learning. While processing external data is an ingestion point, the risk of indirect prompt injection is negligible due to the nature of the data (images) and the specialized CLI tools used for processing.\n
- Ingestion points: User-provided image directories defined in training, inference, and distillation specifications (e.g.,
dataset.train_dataset.images_dir).\n - Boundary markers: Not explicitly used, as the input data is binary image content rather than natural language instructions.\n
- Capability inventory: The skill provides access to the TAO Toolkit CLI (via Docker) to perform model training, export, and inference.\n
- Sanitization: The skill instructions recommend extracting and verifying archives before processing, following standard data handling practices.
Audit Metadata