OpenAI's AI Software Factory Can Bypass Human Review — But Who Audits the Gatekeeper?

OpenAI has built an agentic software factory that allows certain pull requests to skip human review when an automated risk classifier labels them low-risk. While individual components like CI pipelines, specialist review agents, and deployment monitors are each evaluated, the routing decision that removes human oversight largely goes unexamined. Critics point out that risk is not a static property — a change deemed low-risk can become high-risk as dependencies shift, permissions change, or product requirements evolve. The concern is not that any specific failure has been observed, but that the classifier's judgment is treated as a label rather than a testable behavior with measurable consequences. Robust evaluation of such a system would require testing boundary cases and tracking whether bypassed human review correlates with rollbacks, incidents, or production drift.