OpenAI releases GPT-6 Cyber: a model specifically designed to find vulnerabilities. Firefox discovered 81 unknown high-risk vulnerabilities, and 948 Linux kernel vulnerabilities were exploited to create executable privilege escalation tools; ExploitGym achieved a 48.5% success rate, 51% higher than the previous generation. Only available to approved security research clients. AI's strongest spear has also become its strongest shield. Can the defender really outrun the attacker?