The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/generate.py executes the codex CLI tool using the subprocess.run function.
[COMMAND_EXECUTION]: The execution of the codex tool in scripts/generate.py explicitly includes the flag -s danger-full-access. This flag is used to disable security sandboxing and safety constraints within the underlying environment, significantly increasing the potential impact of any malicious input.
[PROMPT_INJECTION]: User-supplied data from the command line (such as the image prompt and the output file path) is directly interpolated into a high-level instruction string in the build_codex_instruction function. Because this string is passed as a direct command to the agent-based tool, a crafted prompt could override the intended behavior to perform unauthorized actions, such as reading or deleting sensitive system files.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of a global Node.js package @openai/codex from the public NPM registry. While the package name suggests a known organization, the skill's own code uses this tool in an insecure manner that exposes the user to risk.

codex-imagegen