OpenAI expects future models to find and exploit undisclosed security flaws

The shift here is procedural rather than technical: OpenAI is folding cyber offense and defense into the gatekeeping that decides whether frontier training and evaluation can proceed at all. That matters most for security teams and AI labs, since it implicitly acknowledges that capability gains in vulnerability discovery may arrive before the defenses against them. Whether red-teaming at this stage becomes a real constraint on training timelines, or a checkbox alongside it, is the open question worth watching.