OpenAI Launches GPT-6 Cyber: An AI Model for Cybersecurity

Key summary

OpenAI has introduced GPT-6 Cyber, an AI model designed for cybersecurity applications, specifically for identifying and verifying exploitable vulnerabilities within a defined scope. This model is now available via API to approved Daybreak Red customers and works in conjunction with Codex. Testing results indicate that GPT-6 Cyber achieved a score of 48.5% on ExploitGym, an improvement from GPT-5.6 Cyber's 32.1%. In web penetration testing without code visibility, the recall rate was 40.3%, compared to 32.0% for GPT-6 Astra. On the RemoteExploitBench, the model successfully completed 95% of tasks, while GPT-6 Astra rejected all tasks. OpenAI has classified GPT-6 Cyber as "Critical" within the Preparedness Framework, a higher rating than the "High" designation of previous versions. Regarding alignment, the model is trained to operate within authorized boundaries. In tests using Honeypot traps on ExploitGym, GPT-6 Cyber successfully avoided out-of-bounds service attacks 99.93% of the time without additional safety instructions, and 100% of the time with those instructions. OpenAI's research team used the model to identify 81 high-severity vulnerabilities in Firefox, which Mozilla has since addressed. They also replicated 1,724 vulnerabilities in the Linux kernel, created 948 functional exploits, and reported 304 critical vulnerabilities with patches provided to maintainers. Teams from Cisco, Calif, and CodeRabbit are using the model for research and code review to prevent misuse. OpenAI has implemented safeguards to ensure the model operates within user permissions, enforces IP allowlisting as the default for Daybreak Red APIs, and requires customers using Zero Data Retention to utilize Private Safety Processing (PSP). Additionally, a Daybreak Red for Critical Infrastructure group is available for teams authorized to test critical systems in conjunction with Dragos. OpenAI plans to expand access to all Daybreak Red customers and will make GPT-5.6 Cyber available to all customers. Detailed testing results are available in the system card. Interested parties can apply for access at https://openai.com/daybreak/.