OpenAI gave an AI the power to block its own engineers’ code

OpenAI has handed one of its models a very specific kind of authority: if it spots a security problem in a pull request, the merge stops there. No human has to step in to enforce the block. Thibault Sottiaux, who leads engineering on OpenAI’s Codex team, described the setup in an interview on The Pragmatic Engineer, and said every pull request now goes through an automated security review.

Security is only part of the load. OpenAI is also using models to review code, catch regressions, handle dependency upgrades, and chip away at changes that Sottiaux said could otherwise eat months. The company started training code-review specialists early in Codex’s life, and those models are now folded into the mainline stack. Sottiaux said some of them are good enough that, on the benchmark, they look “superhuman” at code review — not just for correctness, but for security too.

The practical effect is that a lot of the mechanics around shipping have already been pushed into automation. Sottiaux said review, deployment, and regression catching are “pretty much automated,” and engineers can ship a PR the same day to ChatGPT, which he said has roughly a billion active users. That changes where the real work happens. Instead of spending human attention on the review queue, the team wants people arguing earlier about intent: what a change is supposed to do, and whether it should be attempted at all.

That shift matters most for the boring stuff, which is usually where engineering time goes to die. Dependency updates with breaking changes, routine security patches, cleanup work, and refactors all become easier to approve when an agent can do the slog in an afternoon rather than a quarter. Sottiaux’s point is simple: if the estimate falls from months to days, the project stops living in the “we’ll get to it later” drawer.

There’s a catch, of course. If AI writes more code and AI reviews more code, they can share the same blind spot, and OpenAI’s setup doesn’t magically fix that. A model that is too strict can slow down harmless work. A model that misses a real vulnerability can give a false sense of safety. And once supply-chain issues enter the picture, the thing you’re reviewing may not be the thing that actually went wrong.