OpenAI hands merge-blocking authority to its own security review model

A review model scores a diff. What turns that score into a control is what the repository is configured to do with it, and by Sottiaux's account the configuration is absolute: a flagged security issue blocks the merge without exception ^[4], and no human reviewer is needed to enforce the check ^[3]. Together those give the review model the authority to stop a merge outright, not just the role of scoring it.

The interview covers only a little of what you would need to run the same thing: the false-positive rate, the escalation path when a block is wrong, which model performs the review, and how many pull requests it has stopped ^[14]. In practice, a gate with no exception process still has one, it just is not written down.

The benchmark claim needs the same reading. Sottiaux says that when OpenAI benchmarks these models, "it's like they're superhuman in code review," and that this holds for security as well as correctness ^[5]; OpenAI has started labelling some code-review models that way internally ^[13]. For that to transfer to your repository, several things would have to be true at once. The vulnerability classes in their evaluation set would have to resemble yours. The model would need the same context their engineers gave it: internal frameworks, conventions, threat model. And the human baseline it beat would have to look like your reviewers on your code. The result as stated is a measurement of OpenAI's workload.

The capability path is the part I find genuinely well engineered. Specialized review models were trained early in Codex's development, and those capabilities have since folded into the mainline models ^[6]. Sottiaux's other example points the same way: Codex shipped a /goal command to hold a model on one objective for days or weeks, he calls it a crutch, and newer models do not need it. "You don't need slash goal anymore. You don't need a harness around it" ^[10]. Harnesses get deleted when the model catches up, while a merge policy tends to outlive the model version it was written against, which is why "blocks without exception" is a heavier commitment than the scaffolding around it.

On maintenance, the arithmetic is worth doing. A refactor estimated at two or three months of engineering work is roughly 42 to 63 engineer-days ^[15]. Sottiaux's argument is that agent work can bring such a cleanup down to days ^[12], and that dependency upgrades land in an afternoon when there is a clear changelog and decent documentation ^[11]. The conditional is doing real work in that sentence. Where it holds, the change lands on the approval decision rather than the review queue: shelved cleanups become cheap enough to fund.

Around all of this, Sottiaux describes review, deployment and regression-catching as pretty much automated, with engineers able to ship a pull request the same day to ChatGPT, which he puts at roughly a billion active users ^[8]. His view is that human judgement moves earlier, into a discussion of intent during planning instead of the review queue ^[9].

Adopt the gate and you inherit one line of enforcement config plus a question OpenAI has not answered in public: who signs off when the model is wrong, and what they are allowed to override.