OpenAI launches GPT-6 Cyber to automate vulnerability discovery

A leap in security research

OpenAI officially released GPT-6 Cyber on November 5, 2026, as part of a broader effort to build AI systems for cybersecurity research. The model is designed to help security professionals find vulnerabilities and perform defensive operations more efficiently. It represents a significant jump in performance compared to its predecessor, GPT-5.6 Cyber.

The company measured these gains using ExploitGym, a benchmark that tests if an AI can successfully exploit specific vulnerabilities in a controlled environment. GPT-6 Cyber reached a 48.5% success rate on this benchmark, while the previous model sat at 32.1%. This improvement suggests the model is becoming much more capable at understanding complex software flaws.

Beyond just exploiting known bugs, the model showed a high proficiency in discovering new ones. In tests involving web penetration testing without access to the source code, GPT-6 Cyber identified 40.3% of vulnerabilities. This outperformed the GPT-6 Astra model, which managed to find 32% of the same issues.

Real-world software testing

The model has already been put to work on some of the world's most widely used software components. During an investigation of the Firefox browser, GPT-6 Cyber found 81 previously unknown vulnerabilities that Mozilla classified as high severity. It also provided evidence that 30 other existing issues were actually more severe than they had been originally reported.

The research extended into the Linux kernel, where the model reproduced 1,724 vulnerabilities and created working exploits for 948 of them. In comparison, the previous GPT-5.6 Sol model only successfully exploited 166 vulnerabilities under the same framework. OpenAI shared 304 of these critical Linux findings with maintainers to help them issue patches.

Security teams at major companies like Cisco and CodeRabbit are already using the technology for code reviews and vulnerability research. The model was also used to investigate the Brotli compression library, which is a common component in Node.js. These real-world applications show how the model can accelerate the process of hardening essential digital infrastructure.

Protecting critical infrastructure

OpenAI is also focusing on the security of industrial systems, such as those used in power plants and water treatment facilities. They partnered with Dragos, an industrial cybersecurity company, to test the model in a simulated operational technology environment. During a test on a mock power plant, the model identified assets and found two previously unknown vulnerabilities in engineering software.

The model was able to separate operational technology assets from general IT noise on its own. It discovered flaws in software that manages protective relay data, which could expose breaker trip logic. This type of reconnaissance is exactly what an adversary would look for before attempting to touch physical equipment.

To manage these risks, OpenAI developed specific safeguards for industrial control systems. These protections are intended to prevent the AI from performing unauthorized activities that could affect safety-critical environments. The goal is to give defenders a map of potential attack paths before real adversaries can find them.

Access and safety controls

OpenAI is managing the rollout of these capabilities through its Daybreak cybersecurity program. This program offers different levels of access depending on the user's needs and authorization. Daybreak Blue is for verified defenders doing research, while Daybreak Red provides advanced capabilities for authorized penetration testing and exploit development.

A new category, Daybreak Red for Critical Infrastructure, was created for approved organizations that protect essential services. Additionally, OpenAI launched a $1 billion Daybreak for Frontline Defenders grant program. This initiative will subsidize access to these advanced security tools for government agencies and other resource-constrained organizations.

Safety remains a core component of the launch, with specific monitoring for cybersecurity misalignment and threat detection. In tests, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of cases without specific instructions. When deployment safety instructions were enabled, the model stayed within authorized boundaries in 100% of the evaluated cases.