OpenAI's Codex is safe to use, and in our cross-tool study it posted the lowest high-severity flaw rate of any tool at 9.8%. But that is not a meaningful lead, the gap is within statistical noise, and Codex apps still leak the same things: hardcoded secrets, disabled TLS, weak randomness. Its corpus skews toward ML and tooling, which shifts the mix slightly, but the root cause is identical.
From our scan data
Codex apps had a 9.8% high-severity flaw rate in our scan, the lowest we measured but not a statistically significant difference.
The risks that actually matter
Hardcoded credentials
Codex inlines keys to make integrations run, and those values reach committed source.
Injection-shaped flaws
Because Codex builds more tooling and model servers, its mix leans slightly more toward command injection and disabled TLS than pasted web keys.
Weak randomness
Math.random used where a cryptographic source is needed.
How to secure a OpenAI Codex app
- Add a secrets rule to AGENTS.md so Codex reads keys from the environment.
- Install a gitleaks pre-commit hook.
- Review any shell command construction and TLS settings.
- Scan the repo before publishing.