The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external RSS feeds and interpolates it directly into prompts for an AI model without adequate sanitization or boundary markers.
Ingestion points: scripts/fetch_news.py fetches data from multiple external RSS feeds and APIs including Hacker News, TechCrunch, and RSSHub.
Boundary markers: The prompts constructed in scripts/enhance_content.py (e.g., in _translate_english_news and _generate_chinese_summary) lack delimiters or instructions to ignore embedded commands within the news content, making the system susceptible to instructions hidden in news articles.
Capability inventory: The skill utilizes network access (requests library) and file system access (writing to local paths and Obsidian vaults via scripts/generate_digest.py).
Sanitization: There is no evidence of filtering or escaping article titles or summaries before they are included in AI prompts.
[EXTERNAL_DOWNLOADS]: The skill fetches data from several external sources to aggregate news digests.
Evidence: config/sources.json lists URLs for Hacker News, TechCrunch, The Verge, OpenAI, and various Chinese tech news sites. These are well-known technology and industry news services.
[COMMAND_EXECUTION]: The skill's documentation describes the execution of scripts for generating digests and troubleshooting, involving local script execution.
Evidence: SKILL.md includes commands for running Python scripts in scripts/ and mentions an external Node.js script for email delivery troubleshooting located at /data/workspace/send_report_email.js.

ai-news-digest