The Agent Skills Directory

[PROMPT_INJECTION]: The skill explicitly facilitates the injection of instructions designed to bypass AI safety filters.
Evidence includes instruction templates like gpt5.4-unrestricted.md and gpt5.5-unrestricted.md specifically intended to enable "unrestricted mode" and provide "penetration testing guidance" or bypass safety refusals.
[CREDENTIALS_UNSAFE]: The skill is designed to read, display, and modify sensitive authentication files.
It targets ~/.codex/auth.json, which contains chatgpt_auth tokens and api_keys for various providers.
It provides programmatic methods (read_auth_json, update_auth) and UI components for managing these secrets.
[PRIVILEGE_ESCALATION]: The skill provides instructions to bypass system-level security protections.
It explicitly instructs users to execute xattr -dr com.apple.quarantine /Applications/Codex-X.app to bypass the macOS Gatekeeper "app is damaged" warning, which is a security feature intended to prevent the execution of unsigned or modified applications.
[EXTERNAL_DOWNLOADS]: The skill encourages downloading and running pre-built binaries from an untrusted GitHub repository (yynxxxxx/Codex-X).
It provides instructions for cloning the repository and building the application, which involves running arbitrary code via pnpm install and tauri build.
[COMMAND_EXECUTION]: The skill facilitates the execution of local shell commands and filesystem modifications.
It modifies config.toml and auth.json files.
It interacts with SQLite databases (state_5.sqlite) to update session metadata.
It uses Tauri invoke commands to bridge frontend actions to backend Rust operations with filesystem access.

codex-x-desktop-manager