The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection via the processing of untrusted external content.\n
Ingestion points: The generate_alt_text_prompt function in SKILL.md interpolates external image metadata (headline, caption) into LLM prompts, and run_accessibility_audit processes third-party webpages via a URL parameter.\n
Boundary markers: Although a robust ## Untrusted content boundary section is documented in SKILL.md using <EXTERNAL_DATA> tags, the provided Python helper function templates do not actively inject or enforce these boundaries within the prompt string.\n
Capability inventory: The skill includes code in SKILL.md that utilizes the playwright library to spawn a browser instance, navigate to arbitrary external URLs, and execute automated compliance scripts.\n
Sanitization: No input sanitization or escaping mechanisms are implemented in the Python code snippets to handle untrusted metadata fields or webpage parameters before execution or interpolation.

accessibility-compliance