The Agent Skills Directory

[SAFE]: The skill contains official documentation and boilerplate code for the Hugging Face Gradio library, which is a well-known service. All referenced URLs are for official documentation sites or trusted media hosting.- [DATA_EXFILTRATION]: The skill documents components that can access hardware (webcam, microphone) and local files (gr.File, gr.Image). It also details the gradio predict CLI for remote interaction. These are core features of the UI framework for its intended purpose and do not show signs of malicious exfiltration logic.- [DYNAMIC_EXECUTION]: The documentation for gr.HTML includes an example of using internal JavaScript (js_on_load) to handle UI events (a star rating system). This is a standard extension mechanism for the library and the code provided is benign.- [INDIRECT_PROMPT_INJECTION]: The skill defines UI components that process untrusted data (e.g., gr.Chatbot, gr.Textbox). 1. Ingestion points: User input enters via components like gr.Textbox and gr.Chatbot documented in SKILL.md. 2. Boundary markers: The framework provides properties like sanitize_html (defaulting to True) and render_markdown to control how data is displayed. 3. Capability inventory: The skill allows the agent to define Python functions (fn=) and JavaScript event handlers (js=) to process input. 4. Sanitization: Standard HTML and Markdown sanitization is available and documented in the component signatures.

hugging-face-gradio