OpenAI is running another internal security sprint with Cathedral Systems, using AI to probe and reinforce its own defenses. No timeline or scope metrics are published, but the initiative signals that OpenAI is treating adversarial AI testing as a recurring operational practice, not a one-time audit.
The details buried in the original are worth your time: the choice of Cathedral Systems as a partner, the specific role AI plays in attack simulation versus defense hardening, and what this says about how frontier labs are beginning to institutionalize red-teaming at scale.
If your mental model of AI safety is still 'alignment research plus policy', this sprint is a corrective. Operational security at the infrastructure level is now a parallel track, and the tools doing the attacking and defending are the same class of models OpenAI ships to the public.
[READ ORIGINAL →]