tao-train-grounding-dino
Pass
Audited by Gen Agent Trust Hub on Aug 25, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill consists primarily of documentation and configuration templates (YAML/JSON) for the NVIDIA TAO Toolkit Grounding DINO model.
- [EXTERNAL_DOWNLOADS]: The skill mentions downloading BERT weights (bert-base-uncased) from HuggingFace Hub, which is a standard requirement for this model architecture and a well-known service.
- [COMMAND_EXECUTION]: Instructions provided involve executing standard TAO Toolkit CLI commands (e.g., 'grounding_dino train') within official NVIDIA Docker containers (nvcr.io/nvidia/tao/tao-toolkit), which is the intended use case for this development tool.
- [CREDENTIALS_UNSAFE]: While schemas include an 'encryption_key' field, it is used for standard model protection within the TAO ecosystem and does not contain hardcoded secrets or instructions for unsafe handling.
- [INDIRECT_PROMPT_INJECTION]: The skill processes user-provided captions for open-vocabulary detection. While this constitutes an ingestion surface, the processing is handled by the BERT text encoder within the model architecture, and no exploitable capabilities or command execution paths based on this input were found.
Audit Metadata