The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill exposes the sabx extension install <repo> command, which allows the installation of code from arbitrary GitHub repositories. This provides a mechanism for executing untrusted third-party code within the agent's operating environment.\n- [COMMAND_EXECUTION]: Includes commands to control host power states: sabx queue complete-action shutdown, hibernate, and standby. These allow the agent to terminate the session, put the host to sleep, or shut down the machine entirely.\n- [EXTERNAL_DOWNLOADS]: The primary tool installation is performed via go install github.com/avivsinai/sabx/cmd/sabx@latest, which fetches code from a third-party GitHub repository.\n- [PROMPT_INJECTION]: The skill presents an indirect prompt injection surface by ingesting data from external, potentially attacker-controlled sources.\n
Ingestion points: RSS feed data, NZB file metadata (titles, categories, script names), and log files accessed through commands like sabx rss list, sabx queue list, and sabx logs list.\n
Boundary markers: The instructions do not define delimiters or specific safety instructions to ignore or isolate commands embedded in the processed data.\n
Capability inventory: The agent has access to sensitive capabilities including remote extension installation, system power management, and broad configuration modification.\n
Sanitization: No evidence of input validation or content sanitization is provided for data retrieved from remote Usenet or RSS sources.\n- [CREDENTIALS_UNSAFE]: The skill mentions a configuration option SABX_ALLOW_INSECURE_STORE which allows the application to store credentials in a plain file instead of the OS keyring. It also supports accepting API keys via environment variables (SABX_API_KEY) and command-line arguments, which can lead to credential exposure in shell history or process listings.

sabx