tao-train-dinov3
Pass
Audited by Gen Agent Trust Hub on Sep 23, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external image data for training and inference, which represents a standard attack surface for indirect prompt injection. This is an inherent requirement of the training workflow and is documented as a minor concern.
- Ingestion points: Image directories defined in
dataset.train_dataset.images_diranddataset.test_dataset.images_diracrossSKILL.mdand spec templates. - Boundary markers: None identified, which is typical for machine learning tasks involving binary image datasets.
- Capability inventory: Execution of training, inference, and conversion commands via
docker runas specified inSKILL.mdandreferences/skill_info.yaml. - Sanitization: No sanitization is performed on the raw image inputs, which is standard practice for self-supervised learning on pixel data.
Audit Metadata