OpenAI Releases GPT-6 Cyber Model with Advanced Security Capabilities

OpenAI announced the release of GPT-6 Cyber, a specialized cybersecurity model that achieved a 51% improvement on the ExploitGym benchmark compared to its predecessor. The model successfully reproduced numerous Linux kernel vulnerabilities and discovered 81 new high-risk vulnerabilities in the Firefox browser. This development highlights the accelerating role of frontier AI models in both identifying software flaws and developing automated exploits at scale. It emphasizes the critical window where AI is increasingly utilized to proactively secure complex software ecosystems before malicious actors can exploit them. GPT-6 Cyber was classified with a 'Critical' cybersecurity risk level under OpenAI's preparedness framework and was initially made available exclusively to approved Daybreak Red customers. The model incorporates strict safety mechanisms such as abuse monitoring, sandboxing, and human oversight to prevent unauthorized attacks.