The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent on how to process untrusted external data (such as user-provided PDFs, Office documents, and web URLs) and convert it into Markdown for further LLM processing. This creates a surface for indirect prompt injection attacks where instructions hidden in documents could influence agent logic.
Ingestion points: The convert(), convert_uri(), and convert_local() methods defined in references/api/python-api.md and illustrated in examples/basic/common-scenarios.md allow the agent to ingest arbitrary file and URL content.
Boundary markers: The skill includes specific security warnings in SKILL.md (Section 6) and examples/basic/common-scenarios.md (Scenario 2) advising on input sanitization and SSRF prevention, though these are human-readable instructions rather than technical delimiters.
Capability inventory: The documented functionality includes reading local files and making network requests via the library's integrated requests session.
Sanitization: While the skill provides high-level mitigation guidance (e.g., using convert_local to avoid network risks), it does not implement automated sanitization or filtering of the content extracted from documents before it is processed by the agent.

markitdown-awesome