The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides extensive documentation for the --dangerously-skip-permissions flag. This flag is designed to auto-approve all tool permission requests without human interaction, effectively bypassing the security guardrails that require user review for potentially dangerous actions like file modifications or command execution.
[REMOTE_CODE_EXECUTION]: The skill documents the agy plugin install and agy mcp add subcommands, which facilitate the download and execution of external code from arbitrary sources, including GitHub repositories (e.g., github.com/owner/repo/plugins/analyzer@main) and remote HTTP/SSE endpoints. These features allow the agent to ingest and execute third-party extensions that have not been verified.
[PERSISTENCE]: The skill documents the agy remote-control start command, which registers and starts a background daemon service with the operating system's service manager (systemd, launchd, or background process). This mechanism allows the agy service to remain active and reachable across user logouts and reboots, providing a persistence vector.
[INDIRECT_PROMPT_INJECTION]: The skill enables an attack surface where the agent processes untrusted external data (via view_file or read_url_content) and has powerful tool access (run_command, write_file, plugin install).
Ingestion points: Files in the workspace, external URLs, and MCP server outputs.
Boundary markers: None explicitly defined in the provided instructions to protect against embedded malicious commands.
Capability inventory: Full file system write access, arbitrary command execution (optionally auto-approved), and plugin management.
Sanitization: No specific instructions for sanitizing or escaping untrusted external content before processing.

antigravity-cli