tao-run-deft-od-aoi
Pass
Audited by Gen Agent Trust Hub on Sep 21, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill utilizes pre-trained model weights from trusted sources, specifically Google's SigLIP model for image embeddings and NVIDIA's Cosmos3-Nano for synthetic defect generation. These references are standard for the deep learning tasks performed by the skill.\n- [COMMAND_EXECUTION]: The
scripts/prepare_deft_od_aoi_sources.pyscript executes theannotations mergecommand to process dataset shards. This is a functional requirement of the TAO Data Services environment. The command is executed with a fixed argument list and internally generated paths, preventing injection risks.\n- [INDIRECT_PROMPT_INJECTION]: The skill possesses a data ingestion surface for external COCO JSON files.\n - Ingestion points: Found in
scripts/prepare_deft_od_aoi_sources.pyandscripts/admit_deft_od_aoi_coco.pywhich load external annotations.\n - Boundary markers: Absent; the skill relies on structured JSON parsing without explicit 'ignore' delimiters for text-based metadata fields.\n
- Capability inventory: Includes subprocess execution of TAO commands and extensive file system write operations across several preparation and admission scripts.\n
- Sanitization: Present; the scripts implement strict validation for bounding box coordinates, category IDs, and image path resolutions to ensure they conform to expected schemas.\n- [DYNAMIC_EXECUTION]: The workflow generates YAML and JSON configuration files used to drive the training and evaluation phases. These files are constructed from predefined templates and validated state, ensuring that dynamic content remains within safe, expected boundaries.
Audit Metadata