Every pull request submitted by an OpenAI engineer now undergoes an automated security review performed by an AI model that has the authority to stop code from being merged if it finds a vulnerability. Thibault Sottiaux, engineering lead of OpenAI's Codex team, described the system in an interview on The Pragmatic Engineer, emphasizing that the security check is mandatory and does not require a human reviewer to enforce it.
Beyond security, OpenAI is using its models to review code for correctness, catch regressions, and handle dependency upgrades. Sottiaux said these tasks could previously have taken months and noted that some of the company's code-review models have been benchmarked as “superhuman.”
“When we benchmark them, it’s like they’re superhuman in code review,” Sottiaux said. “This is not just true for correctness. This is also true for security.”
As AI takes over more of the mechanics of reviewing code, Sottiaux sees the human role moving earlier in the process. OpenAI's review, deployment, and regression-catching processes are already “pretty much automated,” engineers can ship a pull request the same day to ChatGPT, which Sottiaux said serves roughly a billion active users.
“Really what we see, and I see, is there’s this sort of discussion around the intent that takes place around the pull request,” Sottiaux said. “It’s like, what are you even trying to do? And is that the right thing to attempt to do?”
Sottiaux argued that this discussion needs to happen earlier, back in the planning phase rather than waiting for the review queue. Engineers still need to agree on the goal and assess a proposed change, but the review burden can be passed to AI without removing humans from the loop — it simply shifts the gut-check to before anyone opens a pull request.
The article also notes that basic maintenance tasks, such as updating third-party libraries with breaking changes or routine security patches, can be handled by AI agents as long as there is a clear changelog and decent documentation. According to Sottiaux, agents can knock out those updates in an afternoon. The same logic could apply to larger refactoring jobs that teams might postpone due to estimated months of engineering work.