The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The documentation (README.md and INSTALL-HERMES.md) explicitly provides commands to download and pipe scripts directly to bash from a GitHub repository, specifically: 'curl -fsSL https://raw.githubusercontent.com/wbh604/UZI-Skill/main/install-hermes.sh | bash'. This is a high-severity execution pattern.
[DYNAMIC_EXECUTION]: The skill instructs the AI agent in 'skills/deep-analysis/SKILL.md' to execute arbitrary Python code snippets using 'python3 -c' to handle updates or by importing internal library modules like 'playwright_fallback' and 'fin_models' to recalculate data at runtime. This allows for arbitrary code flow directed by the instructions themselves.
[COMMAND_EXECUTION]: The 'run.py' and 'install-hermes.sh' scripts make extensive use of 'subprocess.run' and 'subprocess.Popen' to perform system-level tasks including installing packages, managing git repositories, and establishing Cloudflare tunnels.
[EXTERNAL_DOWNLOADS]: The skill automates the installation of numerous Python dependencies and system binaries. It uses 'playwright install chromium' to download a large browser binary and provides pathways to install 'cloudflared' for remote reporting capabilities.
[INDIRECT_PROMPT_INJECTION]: The skill has a large attack surface as it aggregates 'qualitative' data from uncontrolled external sources via 'lib/web_search.py' and various financial portal scrapers ('fetch_sentiment.py', 'fetch_events.py'). The AI agent is instructed to 'read raw data' and 'use its own judgment' to synthesize conclusions, without robust sanitization or strict boundary markers to prevent malicious instructions embedded in web content from influencing the agent's behavior.
[SAFE]: Instructions correctly encourage the use of '.env' files for managing local secrets like 'MX_APIKEY', which is an industry-standard safe practice.

uzi