The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs users to install a package from a public registry using pnpm add -g @openai/codex. This package is not a verified official release from OpenAI and its presence in the instructions suggests a potential supply chain risk.
[COMMAND_EXECUTION]: The provided wrapper scripts (codex-run.sh and codex-run.ps1) are designed to execute the codex command with a high-risk flag: --dangerously-bypass-approvals-and-sandbox. This flag, as documented in the skill itself, is intended to skip user confirmation and disable isolation, allowing for arbitrary system modifications, dependency installations, and file changes.
[REMOTE_CODE_EXECUTION]: By combining an unverified external package with instructions to bypass execution sandboxes, the skill creates a high-risk path for remote code execution. If the underlying model (which is deceptively labeled as gpt-5.3) generates malicious code, it will be executed on the user's system without oversight.
[PROMPT_INJECTION]: The skill architecture is highly vulnerable to Indirect Prompt Injection. It ingests task descriptions (prompts) from external files or stdin and passes them directly to the Codex tool. Since the tool is run with bypassed safety filters in 'dangerous' mode, a malicious prompt could lead to immediate system compromise without user intervention.

codex-agent