What's this about?
OpenAI describes a continuous security operation in which agents inventory systems, find findings, reproduce them, identify owners, prepare patches and verify deployed fixes. A 250-person sprint closed 53 urgent or high-priority issues on its first day.
What exactly is a “Defense Factory”?
It is a reusable loop rather than a one-off scanner. Each finding carries system context and evidence through triage, an isolated reproduction environment, human-reviewed remediation and an independent final retest.
Why is this worth noting?
OpenAI reports 90.6% accepted ownership assignments, 37% duplicate findings, 19.5% runtime reproduction and a 0.81% false-positive rate after validation. All sprint patches were generated with Codex, but consequential changes retained human review; the metrics are internal and not independently audited.