The Agent Skills Directory

[PROMPT_INJECTION]: The skill is designed to ingest and convert untrusted content from external URLs and local files, which presents a surface for indirect prompt injection attacks.
Ingestion points: The scripts/render.js script fetches arbitrary content via the --url argument and reads local documents via the --file argument.
Boundary markers: The script includes a dedicated printSafeOutput function that wraps the converted Markdown in explicit 'START OF UNTRUSTED CONTENT' and 'END OF UNTRUSTED CONTENT' markers, accompanied by a warning for the agent to treat the content as data.
Capability inventory: The script uses the fetch API for network operations and fs.readFileSync for local file access.
Sanitization: The tool relies on boundary markers and agent instruction rather than semantic content sanitization.
[DATA_EXFILTRATION]: The skill transmits data from local files or web sources to Cloudflare's infrastructure for processing.
Evidence: scripts/render.js sends content via POST requests to https://api.cloudflare.com/. While Cloudflare is a well-known service and this behavior is essential to the skill's primary function, it involves sending potentially sensitive user data to a third-party API.

cloudflare-tomarkdown