The Agent Skills Directory

[PROMPT_INJECTION]: The skill is designed to process user-provided datasets (images and captions) for model training and inference, creating an indirect prompt injection surface.
Ingestion points: The skill ingests external data from image directories, caption folders, and shard lists specified in the configuration (e.g., dataset.train.datasets).
Boundary markers: Data paths are handled via structured YAML and JSON configuration files, but instructions do not mandate specific isolation or sanitization for potential instructions embedded in the captions.
Capability inventory: The agent utilizes the Bash tool to execute clip CLI commands and can write output to a user-defined results directory.
Sanitization: There is no explicit sanitization logic described for filtering or validating the textual content of captions before they are processed by the training or inference containers.
[EXTERNAL_DOWNLOADS]: The skill documentation correctly references official NVIDIA container images hosted on the nvcr.io registry (e.g., nvcr.io/nvidia/tao/tao-toolkit:6.26.3-deploy). These references target well-known services associated with the skill's author and are documented neutrally.
[SAFE]: The skill correctly identifies a security consideration regarding PyTorch 2.6 checkpoint loading (TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1) and explicitly instructs the user not to enable this override for untrusted checkpoints, following established security best practices.

tao-finetune-clip