The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided ComfyUI workflow files (JSON) and PNG images, which is a potential attack surface. However, the included scripts (extract_png_workflow.py, inventory_workflow_models.py, apply_model_metadata.py) implement rigorous validation, including JSON depth limits, strict URL allowlisting (enforcing HTTPS and specific trusted domains like Hugging Face), and recursive JSON pointer validation. The instructions also explicitly guide the agent to reject unsafe URLs and avoid leaking credentials.
[COMMAND_EXECUTION]: The skill uses local Python scripts invoked via Bash as allowed tools. These scripts are self-contained and focus on data processing. Unit tests use subprocess.run() to verify CLI functionality, which is a standard and safe testing practice in this context.
[CREDENTIALS_UNSAFE]: The documentation mentions default credentials for a template (FileBrowser admin/adminadmin12). These are documented defaults for the container images being described, not secrets belonging to the user or the skill environment. The skill instructions reinforce that users should rotate these defaults.

runpod-templates