tao-train-oneformer
Pass
Audited by Gen Agent Trust Hub on Sep 22, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external dataset files and executes training commands within a containerized environment.
- Ingestion points: Dataset images, annotations, and label maps are read from user-specified local or S3 paths as defined in references/skill_info.yaml.
- Boundary markers: Absent; the skill assumes integrity of the processed segmentation data.
- Capability inventory: Shell commands are executed via the 'oneformer' CLI within Docker containers (references/skill_info.yaml).
- Sanitization: Absent; the skill relies on the underlying TAO Toolkit to parse standard segmentation formats.
- [EXTERNAL_DOWNLOADS]: Fetches a pretrained backbone model file from a well-known research repository on GitHub.
- Evidence: SKILL.md references https://github.com/SwinTransformer/storage/releases/download/v1.0.8/swin_tiny_patch4_window7_224_22k.pth.
Audit Metadata