5.2 Codex Security: Catch Bugs Before Ship

The more autonomous your setup gets, the more a real security pass pays for itself — and the more expensive it is to skip. Codex Security is worth treating as non-negotiable the moment you're running Codex on anything beyond a throwaway project.

Codex Security is an application-security agent, built from an internal project called Aardvark, specifically trained to identify and propose fixes for real vulnerabilities in your codebase — not a generic linter flagging style issues, an agent reasoning about actual exploitable weaknesses the way a security-focused engineer would. It's had genuine real-world validation too: security researchers have used it to surface previously unknown vulnerabilities in widely-used open-source projects during legitimate research, which is a meaningfully higher bar than "passes a demo."

Beyond the dedicated security agent, the same discipline from earlier modules still applies and still matters most here. Know your sandbox and approval settings before you walk away from any session touching auth, payments, or user data — loose settings are genuinely great for reviewed, low-risk work and a liability the moment your mindset shifts to "I won't look at this again until it's done" on anything sensitive. Treat MCP servers the same way you'd treat an npm dependency you're about to add to production: check the source, prefer connections your organization controls or you personally trust, and don't connect something purely because it looked useful in a forum thread — the blast radius of a malicious MCP server is larger than a malicious package, because it can feed the agent instructions, not just data.

And run a security-focused review pass on anything before it ships, the same instinct as wanting a second engineer's eyes on a risky change — except this second reviewer is available at 2am and never gets tired on the fortieth PR of the week.