The Agent Skills Directory

[SAFE]: No security threats were identified during the analysis of the skill. The implementation follows security best practices for AI agent extensions.
[COMMAND_EXECUTION]: The skill uses torchrun within scripts/finetune_anomalygennext.sh to initiate model training. This is the intended primary function of the skill. The arguments passed to the shell commands are extracted from a validated configuration file generated by the prepare_recipe action, minimizing the risk of command injection.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data formats such as JSONL (defect_spec.jsonl, validation.jsonl). Risk is mitigated in scripts/prepare_finetune_recipe.py by strictly validating user-provided strings (like dataset_name and job_name) against a restrictive regular expression allowlist (^[A-Za-z0-9][A-Za-z0-9._-]*$). This prevents malicious data from influencing command-line arguments or file paths.
[DYNAMIC_EXECUTION]: The skill utilizes yaml.safe_load() throughout its Python scripts (scripts/prepare_finetune_recipe.py, scripts/validate_nn_improvement.py) and inside the shell wrapper to prevent arbitrary code execution during configuration parsing.
[EXTERNAL_DOWNLOADS]: The skill references an official NVIDIA container image (nvcr.io/nvidia/paidf-anomalygen:1.1.0). This is a trusted source consistent with the skill's authorship by nvidia-tao. The scripts also reference a dinov2-large checkpoint from a fixed path, which is a standard procedure for this machine learning task.
[DATA_EXFILTRATION]: No evidence of unauthorized network requests or data exfiltration was found. File system operations are localized to the dataset, checkpoints, and specified results directories.

tao-finetune-anomalygennext