tao-finetune-clip

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.85). The required runtime workflow for this skill ingests outsider-authored free text from user-provided dataset caption files (e.g., dataset.train.datasets[*].caption_dir / dataset.val.datasets[*].caption_dir and inference.text_file), which are read as text during training/evaluation/inference and can contain arbitrary prompt-like content.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.70). The skill documentation shows it fetches and executes external artifacts at runtime — e.g., pulling the TAO Deploy container nvcr.io/nvidia/tao/tao-toolkit:6.26.3-deploy and loading pretrained weights via HuggingFace/torch.hub (https://huggingface.co / torch.hub), so remote content can be executed or directly affect runtime behavior.