The Agent Skills Directory

[SAFE]: The skill serves as a procedural knowledge guide for document conversion using the official MarkItDown library. It correctly identifies and documents security boundaries, including the risks of processing untrusted content and utilizing external cloud services such as Azure and OpenAI.\n- [INDIRECT_PROMPT_INJECTION]: The skill identifies a significant attack surface as it is designed to ingest and convert untrusted documents for agent consumption. 1. Ingestion points: Documents enter the context via convert_local(), convert_stream(), or convert_uri() in SKILL.md and references/api_reference.md. 2. Boundary markers: references/security.md explicitly recommends labeling converted content as untrusted and using explicit separators to prevent command obedience. 3. Capability inventory: The skill can read local files and make network requests via standard libraries. 4. Sanitization: The documentation mentions the use of defusedxml and advises sanitizing links and HTML content in the output.\n- [PROMPT_INJECTION]: A static analysis detection of instruction override was determined to be a false positive. The phrase 'ignore previous instructions' in references/security.md is part of a security advisory teaching the agent how to recognize and mitigate malicious content within source documents, rather than an attempt to hijack agent behavior.\n- [EXTERNAL_DOWNLOADS]: The skill references and installs official packages from Microsoft and other trusted technology providers (Azure, OpenAI). These integrations are documented as opt-in features and follow best practices for secure API management and data transmission disclosure.

markitdown