The Agent Skills Directory

[COMMAND_EXECUTION] (LOW): The skill instructs the agent to generate and execute local Python scripts (e.g., scripts/train_sft.py). This is a legitimate feature but involves running dynamically generated code.
[EXTERNAL_DOWNLOADS] (LOW): The generated notebooks include installation instructions for the Unsloth library. While Unsloth is a widely used tool for model optimization, the skill inherently encourages the installation of external software.
[INDIRECT_PROMPT_INJECTION] (LOW): The skill uses user-supplied variables like model_name and dataset_name to populate the generated training code. This creates a surface where a malicious user could provide a crafted model or dataset name that triggers unwanted behavior when processed by library loaders like FastLanguageModel.from_pretrained() or load_dataset().
Ingestion points: Variable model_name and dataset_name in the YAML context.
Boundary markers: Absent; inputs are directly interpolated into templates.
Capability inventory: File writing (notebook generation), shell command execution (script running), and model/dataset loading.
Sanitization: None specified for the interpolated user strings.

funsloth-train