OpenAI releases GPT-6 Cyber to find software vulnerabilities

A new tool for finding flaws

OpenAI recently announced GPT-6 Cyber, a model specifically designed to identify holes in software. The company used the model to find 81 previously unknown high-severity vulnerabilities in the Mozilla Firefox web browser. Mozilla has already begun publishing fixes for these specific flaws.

The model's capabilities were tested using an autonomous pipeline. This setup produced 948 working local privilege escalation exploits for Linux. Local privilege escalation is a type of security flaw that allows a user to gain higher-level permissions on a computer system.

Beyond just finding bugs, the model was used to proactively improve software. OpenAI sent 304 critical Linux findings to maintainers along with suggested fixes. This work is part of a broader effort to use AI to secure the underlying infrastructure of the internet.

The dual-use dilemma

The release of GPT-6 Cyber highlights a significant challenge in AI safety. OpenAI classified the model as 'Critical' for cybersecurity under its Preparedness Framework. This rating is one step higher than its predecessor, GPT-5.6 Cyber, indicating that the tool is inherently powerful and potentially dangerous.

Because the model can turn a Linux kernel flaw into a working root exploit, it serves both defenders and attackers. A root exploit gives an attacker complete control over a system. There is no version of a coding model that is better at writing software without also being better at hacking it.

Every laboratory racing to improve AI for coding is simultaneously improving the tools available for cyberattacks. This creates a situation where the same skill set used to defend a company can be used to breach it. The company selling the defensive tool is essentially acknowledging its own risk in its safety paperwork.

Performance and safety tests

OpenAI conducted tests to see if the model could stay within its intended boundaries. When the developers did not use specific safety instructions, GPT-6 Cyber stayed on task in 99.93% of tests. This was an improvement over the 94.49% success rate of the previous GPT-5.6 Cyber model.

When the developers included safety prompts, the model stayed in bounds every single time. However, the researchers noted that lab tests may not perfectly reflect how a model behaves in a messy, real-world customer network. They suggest treating these high scores as vendor figures rather than absolute guarantees.

The model's efficiency was also demonstrated through a specific research pipeline. It reproduced 1,724 real Linux kernel bugs from 2,308 candidates. The process turned 948 of those into confirmed root exploits with patches for roughly $1.2 million in a single run.

Who can access the model

GPT-6 Cyber is not currently available for general public use or purchase. It is initially available through the API and Codex to select customers within OpenAI's Daybreak Red program. Access is currently limited to approved security teams who are part of this specific initiative.

Major security companies have already begun integrating these types of capabilities into their workflows. Reports indicate that firms like Cisco and CodeRabbit have used the model for code review. This allows enterprise customers to scan their own code for vulnerabilities before it is deployed.

The program competes with other industry efforts, such as Anthropic's Project Glasswing. While there is no official price or general availability date yet, the rollout suggests a focus on high-end security partners. These partners will likely be the first to use the tool to defend against AI-driven threats.

The evolving threat landscape

The rise of AI-powered hacking is already being observed in the wild. In May, a group of security experts at Google noticed a criminal organization trying to use AI to find and use a zero-day exploit. A zero-day is a vulnerability that is unknown to the software vendor and has no existing patch.

The attackers targeted a popular open-source, web-based admin tool. The exploit was capable of bypassing two-factor authentication. Google determined that an AI model likely helped the criminals because the Python script they used contained a hallucinated CVSS score.

A CVSS score is a standard way to measure the severity of a security vulnerability. Because the target company patched the flaw before a mass attack began, the damage was contained. This incident serves as a real-world example of how quickly AI can be used to develop functional exploits.