The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the global installation of the package @openai/codex. This package is not an official OpenAI distribution (the standard package is openai), which is a known pattern for dependency confusion or supply chain attacks masquerading as legitimate tools.
[COMMAND_EXECUTION]: The codex_run tool executes shell commands via codex exec. It accepts and passes user-provided parameters including prompt, cwd, and model directly to the system's command-line interface, providing a direct path for arbitrary command execution.
[REMOTE_CODE_EXECUTION]: The skill supports a full-auto mode (--full-auto flag), allowing the unverified CLI tool to execute code or modify the workspace autonomously. Combined with the suspicious external package, this constitutes a high risk of remote code execution.
[PROMPT_INJECTION]: The prompt parameter is used to provide instructions to the underlying CLI tool without visible sanitization or boundary markers. This creates a surface for indirect prompt injection where malicious instructions embedded in user data could be executed by the agent with terminal access.

codex-cli