The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill explicitly instructs the user or agent to install software using a dangerous pipe-to-shell pattern: curl -fsSL https://chatgpt.com/codex/install.sh | sh. This allows for arbitrary code execution from a remote source without verification.
[COMMAND_EXECUTION]: The skill includes instructions to bypass shell aliases using the command prefix (e.g., command codex), ensuring that any security wrappers or logging aliases set by the user are ignored. It also uses the --yolo and --dangerously-bypass-approvals-and-sandbox flags, which explicitly disable safety confirmations and sandboxing within the Codex tool.
[COMMAND_EXECUTION]: The skill provides scripts to dynamically create executable files on the host system, such as generating a wrapper script in ~/.local/bin/codex using printf and chmod +x.
[EXTERNAL_DOWNLOADS]: It references and executes a preflight Ruby script from a hidden directory (~/.codex/skills/agent-scripts/codex-huge-context/scripts/preflight.rb) which may involve external network requests or credential harvesting from the Keychain.
[PROMPT_INJECTION]: The skill contains 'Hard gate' instructions designed to override repository-level instructions and force specific behavior based on the state of environment variables like ANTHROPIC_BASE_URL.

codex-first