The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches XML feed data from external RSS and Atom sources. Evidence: The script uses urllib.request.urlopen in scripts/main.py to retrieve content from user-provided feed URLs. Security measures include restricted URL schemes (HTTP/HTTPS only) to prevent SSRF or local file access, and a 4MB size limit per feed to mitigate resource exhaustion attacks.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from external web feeds that could contain adversarial instructions. Ingestion points: Feed items are parsed and included in the LLM prompt in scripts/main.py. Boundary markers: The prompt explicitly warns the agent that feed content is untrusted text and instructions within it should be ignored. Capability inventory: The agent has access to terminal and file_editor tools. Sanitization: HTML markup and script tags are stripped using regex before the content is passed to the LLM.
[DYNAMIC_EXECUTION]: Dynamically generates a customized Python script during the setup process. Evidence: SKILL.md Step 5 details substituting user-defined feed URLs and topics into a script template. The process includes a validation step using py_compile and safe string serialization via json.dumps to prevent injection into the generated script.

news-digest