tao-generate-od-defects
Pass
Audited by Gen Agent Trust Hub on Sep 17, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses
subprocess.runto invoke model generation (generate.py) and pseudo-labeling (pseudo_label.py) scripts. These scripts are located within the/workspace/paidf-anomalygendirectory of the official container image. - [EXTERNAL_DOWNLOADS]: Execution is tied to a specific, pinned NVIDIA container image (
nvcr.io/nvidia/paidf-anomalygen:1.1.0). It also allows for optional Hugging Face model assets to be cached or downloaded from known repositories. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input data (testcases, recipes, and generation plans). It implements several defensive measures to mitigate injection or data corruption:
- Ingestion points: Reads JSON, JSONL, and YAML files from user-specified input directories.
- Boundary markers: Enforces strict schema validation for COCO and custom metadata formats.
- Capability inventory: Subprocess calls are restricted to internal container scripts; file access is scoped to the provided input and output directories.
- Sanitization: Uses
yaml.safe_loadfor parsing, regex validation for dataset identifiers, and SHA256 hash verification for artifact integrity when a manifest is present.
Audit Metadata