OpenAI launches GPT-6 Cyber to find and verify software security flaws

A new tool for vulnerability research

OpenAI has officially unveiled GPT-6 Cyber, a model specifically engineered to navigate the complexities of software security. Unlike general-purpose AI, this model is built to detect security vulnerabilities and verify if those flaws can actually be exploited in a real-world setting. It serves as a specialized assistant for security researchers who need to find and fix bugs in large, complex codebases.

The model is designed to do more than just point out errors; it also supports the development of security patches. By generating analysis materials and potential fixes, it helps security teams move from discovery to remediation much faster. This proactive approach aims to help defenders identify and close security gaps before malicious actors have the chance to exploit them.

As reported by Herald Business, the model is currently being rolled out to select customers of Daybreak Red. These early users will have access to the model via API and Codex, with plans to expand availability to all Daybreak Red customers in the future. This phased rollout ensures that the tool is used by approved researchers who can manage its capabilities responsibly.

Real-world testing results

The capabilities of GPT-6 Cyber have already been demonstrated through significant research into major software components. OpenAI researchers used the model to uncover 81 previously unknown high-severity security vulnerabilities in the Firefox web browser. For each flaw found, the team generated verification materials and shared them with Mozilla, which then released official fixes based on those findings.

The model also performed extensive research on the Linux kernel, a core component of many operating systems. It reproduced a total of 1,724 vulnerabilities and successfully developed working exploits for 948 of those issues. Following established community guidelines, OpenAI reported 304 critical findings to maintainers along with potential patch candidates while continuing to work on the remaining vulnerabilities.

These results highlight the model's ability to handle high-volume code analysis. By identifying thousands of issues across different platforms, the model shows it can scale security auditing in ways that are difficult for human teams to achieve alone. The focus remains on providing actionable data to maintainers to improve the overall security of the software ecosystem.

Benchmark performance gains

GPT-6 Cyber shows measurable improvements over its predecessor, GPT-5.6 Cyber, across several industry benchmarks. On ExploitGym, a platform that evaluates whether a vulnerability can be exploited in practice, the new model reached a success rate of 48.5%. This represents a 16 percentage point increase from the 32.1% success rate recorded by the previous version.

The model also demonstrated high proficiency in complex environments. In an initial evaluation on RemoteExploitBench, which tests security tasks on servers simulating real-world conditions, GPT-6 Cyber completed 95% of the assigned challenges. These metrics suggest that the model is becoming increasingly adept at navigating the nuances of modern cybersecurity infrastructure.

Furthermore, the model showed improved reliability in staying within its designated boundaries. On the ExploitGym Honeypot benchmark, which measures whether an AI attacks services outside its authorized scope, GPT-6 Cyber refrained from out-of-scope attacks in 99.93% of cases. This is a notable improvement over the 94.49% rate achieved by the older model, even when deployment safety prompts were not used.

Safety and scope constraints

OpenAI has implemented specific training to ensure that GPT-6 Cyber remains a helpful tool rather than a risk. The model received specialized training to adhere to approved targets and boundaries, which helps it report results accurately. This is a critical safeguard designed to reduce the risk of a cybersecurity AI agent accessing systems that it was not specifically authorized to examine.

To further protect users and infrastructure, OpenAI reinforced the model with several layers of monitoring. These systems verify that the model's actions remain strictly within the bounds of what a user has approved. The model is specifically designed to block any requests that could potentially harm critical infrastructure or essential public services while still supporting legitimate research.

By balancing high-powered exploit capabilities with strict scope controls, OpenAI aims to provide a tool for the