OpenAI released GPT-6 Cyber, a model built to find and exploit security flaws, to a small group of vetted customers on Thursday, rating it "Critical" in cybersecurity under its safety framework.