The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements various data ingestion points that handle untrusted external input.\n
Evidence: Components such as gr.Textbox, gr.File, gr.Chatbot, and gr.Audio are designed to receive data from users or external sources, which could contain malicious instructions designed to influence the agent's behavior.\n
Sanitization: The Chatbot and Markdown components include a sanitize_html parameter (enabled by default) to mitigate basic injection attacks.\n- [DYNAMIC_EXECUTION]: The gr.HTML component supports custom templates and client-side script execution.\n
Evidence: The StarRating example demonstrates the use of html_template, css_template, and js_on_load to build interactive components. This functionality allows for the dynamic execution of JavaScript in the user's browser context.\n- [COMMAND_EXECUTION]: The skill documents the use of the gradio CLI for interacting with Gradio applications.\n
Evidence: The gradio info and gradio predict commands allow the agent to discover endpoints and execute remote predictions. These commands interact with the host system's shell and network.\n- [DATA_EXFILTRATION]: The CLI documentation describes capabilities for sending local data to remote endpoints.\n
Evidence: The gradio predict command includes a mechanism for file uploads to remote servers using the gradio.FileData metadata type, which represents a potential channel for data transfer to external services.\n- [SAFE]: The skill originates from a trusted repository and includes proactive security guidance for developers.\n
Evidence: A dedicated "Version and security boundaries" section instructs users to keep demos on loopback by default, validate upload types, and avoid exposing secrets or local directories through the UI.

hugging-face-gradio