The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external URLs or local PDF files. This creates a vulnerability surface where malicious instructions embedded in the source documents could influence the agent's behavior during the translation or slide generation phases.
Ingestion points: The web_fetcher.py script retrieves external content from user-provided URLs or PDF paths.
Boundary markers: The instructions do not specify explicit delimiters or "ignore embedded instructions" warnings when the fetched content is passed to the agent for processing.
Capability inventory: The skill has access to network operations (urllib.request), file system writes, and local script execution (subprocess.run).
Sanitization: The web_fetcher.py script performs basic cleaning of Markdown residues and SVG data but does not sanitize natural language instructions from the input.
[COMMAND_EXECUTION]: The skill requires the agent to execute local Python scripts (web_fetcher.py, generate_slides.py, and build.py) to perform its core tasks.
The generate_slides.py script uses subprocess.run to call the codex CLI tool. The analysis of the script shows it uses a list for command arguments rather than a shell string, which is a safer practice.
The web_fetcher.py script includes an optional --insecure flag that allows the agent to bypass SSL certificate verification (ssl._create_unverified_context) when fetching content, which could expose the session to man-in-the-middle attacks.
[EXTERNAL_DOWNLOADS]: The build.py script generates a preview.html file that loads the marked.js Markdown rendering library from the jsDelivr CDN at runtime.

doc-reader