OpenAI has unveiled GPT-6 Cyber, a significant advancement in security research that demonstrates major improvements in finding and fixing vulnerabilities while staying within defined boundaries. The new model shows a 51% relative improvement over GPT-5.6 Cyber on Exploit Gym, achieving a score of 48.5% compared to 32.1%. On alignment evaluations, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of tests without deployment prompts and 100% with prompts.
The system is already helping secure real-world software, uncovering thousands of exploits in Linux and Firefox, and identifying 81 previously unknown high-severity vulnerabilities in Firefox. Major security teams at Cisco, Code Rabbit, and OpenAI itself are utilizing the technology for vulnerability research. OpenAI has classified GPT-6 Cyber as Critical under its Preparedness Framework, reflecting its advanced capabilities.
The model also supports broader security initiatives through Daybreak, a platform designed to expand advanced security research to organizations and open-source maintainers. This development represents a significant step forward in using frontier models to secure software across various platforms and infrastructures.
Source: OpenAI Blog · Summarized by HeadlinesBriefing