The Agent Skills Directory

[SAFE]: The skill facilitates the installation and use of official packages and source code from Microsoft's public GitHub repository, which is a trusted organization.
[SAFE]: Security best practices for secret management are enforced throughout the documentation. The skill explicitly warns against hardcoding API keys and provides examples using secure environment variable access via os.environ.
[INDIRECT_PROMPT_INJECTION]: The skill processes document formats (PDF, DOCX, PPTX, XLSX) that may contain untrusted data in embedded images. A potential attack surface is identified:
Ingestion points: External documents are ingested via the MarkItDown.convert() method across multiple example scripts.
Boundary markers: The skill implements a distinct output format *[Image OCR] ... [End OCR]* to encapsulate extracted text from images.
Capability inventory: The skill facilitates document reading, network connections to LLM provider APIs (OpenAI, Azure, Google), and file writing through provided CLI wrapper scripts.
Sanitization: The documentation provides a dedicated post-processing mitigation (the strip_ocr_blocks function in examples/advanced/scenarios.md) to prevent the agent from accidentally executing instructions found within document images.

markitdown-ocr