The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to read user-provided Markdown files to extract metadata and context for automated tasks like AI image generation.
Ingestion points: The agent parses the input Markdown file (referenced in SKILL.md) to read titles and opening paragraphs.
Boundary markers: The instructions lack explicit boundary markers or warnings to the agent to disregard instructions that might be embedded in the processed document.
Capability inventory: The skill utilizes the md2pdf.py script which has network access and file system access. The script specifically supports file:// URL schemes in image references, which could be exploited to read local files if the agent is tricked by malicious Markdown content. The agent also uses image generation tools based on the document's content.
Sanitization: No sanitization of the untrusted Markdown content is performed before the agent processes it for decision-making.
[EXTERNAL_DOWNLOADS]: The md2pdf.py script performs network requests to fetch external resources.
Evidence: The script uses urllib.request to download remote images defined in the Markdown input and emoji assets from a public CDN (cdnjs.cloudflare.com). While the CDN is a well-known service, the ability to fetch images from any URL provided in the input Markdown constitutes a wide external download surface.

lov-any2pdf