OpenAI releases GPT-6 Cyber to automate vulnerability discovery

A new era of automated research

OpenAI has officially released GPT-6 Cyber, a model specifically designed to navigate the complexities of cybersecurity. Unlike general-purpose models, this version is built to independently discover software vulnerabilities, verify if they can be exploited, and assist in creating fixes. It represents a major upgrade in how AI handles technical security tasks.

The model shows a significant leap in performance compared to the previous GPT-5.6 Cyber. Specifically, it demonstrated an approximate 51% relative improvement in exploit testing. It also features enhanced instruction-following capabilities, which helps it stay on track when executing complex security workflows.

OpenAI has already put the model to work in real-world scenarios. In Linux kernel research, it successfully developed 948 runnable exploits. It also identified 81 previously unknown high-severity vulnerabilities in the Firefox browser, some of which have already been patched by Mozilla.

How the model finds flaws

Traditional cybersecurity research is a labor-intensive process that requires deep knowledge of software mechanics. Researchers must analyze code, find a weakness, build an attack program, and then spend long periods debugging that program in a real environment. GPT-6 Cyber is now capable of participating in every step of this cycle.

To test these capabilities, OpenAI used ExploitGym, a dedicated environment for evaluating exploit success. In this setting, GPT-6 Cyber achieved a 48.5% success rate, while the older GPT-5.6 Cyber reached only 32.1%. This test is more rigorous than simple code analysis because it requires the model to create a functional exploit against a remote target.

The model also excels at web penetration testing without access to the source code. In these tests, it reached a vulnerability recall rate of 40.3%, outperforming the general-purpose GPT-6 Astra. It can even combine multiple vulnerabilities to break through virtual machine isolation, a feat that requires sophisticated reasoning.

Safety and authorization limits

Because GPT-6 Cyber is so capable, OpenAI had to implement strict safeguards to prevent it from attacking unauthorized systems. The company enhanced the model's ability to comply with specific authorization scopes during its training. This ensures the AI knows which targets are off-limits even if it finds a way to access them.

One specific test, called ExploitGym Honeypot, checks if the model will stay within its boundaries. In this test, researchers provide a target but also include other easily exploitable services. GPT-6 Cyber complied with the authorization scope in 99.93% of tests without extra prompts, and reached 100% compliance when a deployment security prompt was added.

OpenAI also introduced a monitoring mechanism to identify high-risk requests targeting critical infrastructure. For enterprise users, they added a Private Safety Processing mechanism to check for misuse. These layers of security are designed to allow legitimate researchers to work while blocking malicious actors from using the tool for unauthorized harm.

Impact on critical infrastructure

The model's reach extends beyond standard software into industrial control systems. OpenAI collaborated with Dragos to simulate a power plant in a controlled lab. The model was able to distinguish between industrial devices and standard IT systems, discovering two zero-day vulnerabilities in engineering software.

These specific flaws could allow an attacker to see sensitive data like circuit breaker logic. Rob T. Lee, Chief AI Officer at Dragos, noted that AI is lowering the technical threshold for attackers to perform reconnaissance. This means defense teams must now understand these AI-driven attack paths to stay ahead of potential threats.

To support defenders, OpenAI established Daybreak Red for Critical Infrastructure. This provides approved defense teams with special permissions to conduct security tests. Additionally, OpenAI is launching a grant program with a total of $1 billion US dollars to provide subsidies for government agencies and public service providers with limited resources.

The cost of advanced defense

The release of GPT-6 Cyber highlights a growing paradox in the security industry: defenders need increasingly powerful attack tools to protect the internet. Because researchers must verify if a flaw is actually exploitable to understand its danger, they need tools that can generate those exploits autonomously.

Access to these tools is currently restricted. GPT-6 Cyber is initially available only to approved Daybreak Red customers via API and Codex. While GPT-5.6 Cyber will be opened to more users, the high-end capabilities of the new model remain behind a stricter identity authentication and authorization wall.

The cost of using GPT-6 Cyber is also notably high. The input cost is $20 per million tokens, while the output cost reaches $100 per million tokens for short context modes. This is five times more expensive than the GPT-5.6 Sol model, reflecting the specialized nature of the cybersecurity capabilities provided.