The Agent Skills Directory

[Dynamic Execution] (MEDIUM): The skill implements and documents a plugin architecture. Specifically, MarkItDown can be initialized with enable_plugins=True (supported in batch_convert.py), which enables the dynamic discovery and loading of third-party Python packages via entry points. The documentation explicitly encourages users to find and install external plugins from unverified GitHub tags.
[Indirect Prompt Injection] (LOW): The skill is designed to convert untrusted external documents (PDF, Word, etc.) into Markdown specifically for LLM consumption, creating a significant attack surface.
Ingestion points: Content is ingested from local files in scripts/batch_convert.py, scripts/convert_literature.py, and scripts/convert_with_ai.py.
Boundary markers: Absent. The scripts do not use XML-style delimiters or 'ignore' instructions when interpolating converted content into final outputs.
Capability inventory: Scripts have the capability to write to the local filesystem and perform network operations to AI providers (OpenRouter).
Sanitization: Absent. Data is processed and outputted without escaping or structural validation of the converted text.
[Persistence Mechanisms] (LOW): Multiple documentation files (INSTALLATION_GUIDE.md, OPENROUTER_INTEGRATION.md) instruct users to persist sensitive API keys by appending export commands to shell profiles like ~/.bashrc and ~/.zshrc. While common for setup, this facilitates the persistence of plaintext credentials in shell configuration files.
[External Downloads] (SAFE): Installation instructions primarily reference the official Microsoft GitHub repository and PyPI, which are considered trusted sources under the defined security policy.

markitdown