The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits a significant surface for indirect prompt injection through the ingestion of untrusted external content. Evidence: Ingestion points include Reddit threads, X posts, Hacker News stories, and YouTube transcripts collected via various scripts and aggregated in scripts/last30days.py. Boundary markers are absent in the SKILL.md instructions, which do not contain delimiters or directions to ignore commands within research data. Capability inventory: The agent has access to powerful tools including Bash, WebSearch, Read, and Write, providing a significant surface for exploitation if an injection occurs. Sanitization: Gathered content is presented to the agent without sanitization or filtering of potential instructions.
[COMMAND_EXECUTION]: Local script and subprocess management. The skill uses the Bash tool to run scripts/last30days.py. This script further executes several external utilities and scripts via subprocesses, including yt-dlp for YouTube data and a vendored Node.js client for X.com searches. Additionally, the scripts/sync.sh script uses sed to dynamically modify skill metadata files during deployment.
[CREDENTIALS_UNSAFE]: Access to sensitive personal data and local storage of secrets. The research engine is designed to extract session cookies from browser profile databases (Chrome, Safari, Firefox) to facilitate authentication with X.com. While this is a core component of its 'free X search' feature, it involve accessing sensitive credential storage. The skill also manages and reads multiple API keys stored in a plain-text configuration file at ~/.config/last30days/.env.

last30days