The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a workflow that ingests model documentation, README files, and example scripts from external sources (Hugging Face, GitHub) to generate training code. This creates an attack surface where malicious content in an untrusted model repository could potentially influence the generated scripts or configuration. The skill manages this surface through a structured multi-step workflow and human-review gates.
[DYNAMIC_EXECUTION]: To address ecosystem incompatibilities, the skill uses a compatibility registry in references/compat-workarounds.md that evaluates Python expressions against model configurations using the eval() function. These expressions are statically defined within the skill's internal reference files, reducing the risk of arbitrary code execution from external input.
[EXTERNAL_DOWNLOADS]: The skill retrieves model weights, datasets, and training recipes from well-known and established services including Hugging Face (huggingface.co) and GitHub (github.com). It also pulls official container images from NVIDIA's NGC registry (nvcr.io). These operations are fundamental to the skill's purpose and are conducted with established industry sources.
[COMMAND_EXECUTION]: The skill orchestrates the training environment using Docker commands (docker build, docker run). It utilizes specific security-relevant flags such as --gpus all, --user for appropriate file ownership, and bind mounts for data management. These are standard high-capability operations required for GPU-accelerated machine learning tasks.

tao-finetune-huggingface-model