Four layers combine detection, model-based review and investigations; alerts can prompt human intervention. OpenAI acknowledges that false positives limit on-call coverage and that urgent paging is disabled for internal Codex traffic.
METR and Redwood tested an earlier version in July and August. On 29 October METR clarified that, according to OpenAI, most vulnerabilities were patched with one fix still underway. It has not assessed those claims or established the expanded system's coverage.