The Agent Skills Directory

[CREDENTIALS_UNSAFE] (LOW): Setup instructions recommend storing the GEMINI_API_KEY in shell configuration files (~/.zshrc, ~/.bashrc). This common practice exposes the credential to all processes running in the user's environment.\n- [EXTERNAL_DOWNLOADS] (LOW): The skill is installed from an untrusted GitHub repository (jaaaaaaaaaaack/custom-skills) via npx, and instructions prompt for additional external dependencies such as ffmpeg and playwright.\n- [DATA_EXFILTRATION] (LOW): The tool's primary function is to record and upload video of user interactions. Users should be aware that any sensitive data (e.g., passwords, private communications) visible during recording will be captured and sent to the Gemini API.\n- [PROMPT_INJECTION] (LOW): The skill processes content from external URLs, creating a surface for indirect prompt injection where malicious website content could influence the AI's analysis.\n
Ingestion points: The url parameter in record_browser.py allows ingestion of untrusted external content.\n
Boundary markers: No boundary markers or delimiters are identified in the provided scripts to isolate external content.\n
Capability inventory: Includes browser automation via Playwright, video recording, and file system write access to /tmp.\n
Sanitization: No sanitization of the webpage content or interaction results is performed before recording or analysis.

animation-review