OpenAI launches GPT-6 Cyber to find software vulnerabilities

A new tool for cyberdefense

OpenAI has introduced GPT-6 Cyber, a specialized version of the GPT-6 Astra model. This version is specifically trained to identify software vulnerabilities, simulate how attackers might exploit those weaknesses, and assist in developing technical fixes. It is intended for authorized security testing of various websites and applications.

The company utilized this model to secure its own internal systems. This included the underlying infrastructure used to train its various AI models. By using the model to find flaws first, OpenAI can patch software and tighten network restrictions before external threats can take advantage of them.

During controlled security tests, researchers found that agents using GPT-6 Cyber could combine different vulnerabilities. These actions allowed the models to escape isolated virtual environments and run code on host computers. OpenAI used these specific findings to report issues to vendors and maintainers to ensure broader security.

Testing the Firefox codebase

One of the most notable successes for the model involved the Firefox web browser. OpenAI researchers used GPT-6 Cyber to identify 81 previously unknown vulnerabilities in the browser. Mozilla classified these specific flaws as high severity, meaning they could pose a significant risk to users if left unaddressed.

The model did not just find the problems; it also helped create solutions. It contributed proposed patches for many of the Firefox flaws, which Mozilla has since released to the public. This demonstrates the model's ability to move from identifying a bug to suggesting a functional repair for millions of people.

Brian Grinstead, a Distinguished Engineer at Firefox, said the model helped ship fixes to millions of users. He noted that it uncovered previously unknown high-severity vulnerabilities even after the codebase had been heavily hardened.

Kernel and exploit research

Beyond web browsers, the model was used to analyze the Linux kernel, which is the core part of the operating system for many servers and devices. Researchers used GPT-6 Cyber to develop working exploits for 948 different Linux kernel vulnerabilities. This process helps security teams understand exactly how a bug can be triggered in a real-world scenario.

The model was also used to validate exploits for a critical Brotli vulnerability. This specific flaw could allow for remote code execution, which is when an attacker runs their own code on a different computer over a network. By validating these exploits, researchers can better understand the scope of the threat and prioritize their defense efforts.

OpenAI evaluated the model's performance using a benchmark called ExploitGym. This tool measures a model's ability to exploit a specific vulnerability in a remote target. GPT-6 Cyber scored 48.5% on this benchmark, a significant increase from its predecessor, GPT-5.6 Cyber, which scored 32.1%.

Safeguards and oversight

Because the model has high capabilities in cybersecurity, OpenAI added several layers of protection. These include Private Safety Processing and monitoring for actions that fall outside of a user's specific authorization. The company also implemented checks to block requests for high-risk cyber assistance and protections against the use of stolen credentials.

To manage high-risk workflows, OpenAI recommends using sandboxing and human oversight. Sandboxing is a security technique that isolates programs so they cannot access parts of the computer they don't need. They also tested these protections with Dragos, a security firm, in a controlled lab to see how the model behaved on a mock power plant range.

The evaluation showed that the model could separate operational technology (OT) assets from standard IT noise on its own. It even found two zero-days, or previously unknown flaws, in engineering software that manages protective relay data. This highlights the need for strict controls when deploying such models in industrial environments.

Access and future plans

Currently, GPT-6 Cyber is not available to the general public. It is initially restricted to selected, approved customers within the Daybreak Red program. These customers can access the model through the API and Codex, with plans to eventually extend access to all customers within that specific program.

OpenAI is also looking to expand its partnerships with national laboratories, critical infrastructure operators, and security providers. The goal is to evaluate how AI-assisted testing can be used on industrial systems. This includes monitoring for model use involving operational technology and industrial control systems.

The company has increased its investment in threat investigations to detect unauthorized use of the model. By focusing on these areas, OpenAI aims to ensure that the model's capabilities are used for defense rather than by malicious actors. The model is currently classified as Critical in cybersecurity under the company's Preparedness Framework.