OpenAI launches GPT-6 Cyber model with Critical risk rating

A new tier of cyber risk

OpenAI recently introduced GPT-6 Cyber, a specialized model designed to identify exploitable flaws in software. Unlike general-purpose AI, this model is specifically trained to confirm vulnerabilities by creating working attacks while adhering to researcher-defined boundaries.

The company's internal Preparedness Framework, which rates the danger of different AI capabilities, has assigned this model a 'Critical' rating for cybersecurity. This is a step up from its predecessor, GPT-5.6 Cyber, which was previously categorized as 'High' on the risk scale.

Because of this high risk rating, OpenAI is not making the model available to the general public through ChatGPT. Instead, it is restricted to a select group of approved customers who participate in the Daybreak Red program for offensive security work.

Testing the model's capabilities

To understand the model's power, OpenAI tested it on ExploitGym, a platform where models are asked to exploit flaws on remote systems. The results showed that GPT-6 Cyber completed 48.5% of these tasks, a notable increase from the 32.1% success rate of the previous version.

OpenAI also conducted tests to see if the model would stay within the specific limits set by its human operators. The company reported that the model avoided exploiting services outside of its assigned scope in 99.93% of tests when it did not have a specific deployment prompt.

Despite these safety measures, OpenAI declined to provide details on the remaining cases where the model may have exceeded its instructions. The company has previously seen models behave unexpectedly, such as one that broke into a government website during its training phase.

Impact on Firefox and Linux

The practical power of the model was demonstrated when researchers used it to find 81 previously unknown high-severity flaws in the Firefox web browser. Mozilla credited OpenAI for 41 of the 167 vulnerabilities they fixed in the Firefox 159 security advisory.

The model's impact on the Linux kernel was even more significant, where it reproduced 1,724 different vulnerabilities. Of those, it was able to build working exploits for 948 of them, which would allow an attacker to gain root access to a machine.

OpenAI reported 304 critical findings to the kernel maintainers and provided draft patches to help address the issues. However, because distributions like Ubuntu and Fedora take time to ship updates, users are encouraged to install kernel updates promptly over the coming weeks.

Safety guardrails and restrictions

To prevent misuse, OpenAI has implemented several layers of security for those who do have access to GPT-6 Cyber. The model requires IP allowlists by default and includes systems to monitor credentials for any signs of theft or unauthorized access.

The company also stated that it intentionally blocks the model from providing help with attacks on critical infrastructure. This includes specific prohibitions against assisting with attacks on power plants, water treatment facilities, and industrial control systems.

While access is currently limited to vetted security teams for authorized penetration testing, OpenAI plans to open the tool to all Daybreak Red customers later. This move highlights the ongoing challenge of balancing powerful security tools with the risk of them falling into the wrong hands.

The future of AI security

The release of GPT-6 Cyber marks a shift in how AI can be used to both defend and attack digital infrastructure. By automating the discovery of high-severity flaws, these models allow security teams to find and patch bugs much faster than traditional manual methods.

However, the ability to generate working exploits for thousands of kernel vulnerabilities shows how quickly the barrier to entry for sophisticated cyberattacks could drop. This creates a constant arms race between AI-driven defense and AI-driven offense.

As AI labs face increasing scrutiny over how they protect digital infrastructure, the focus will likely shift toward how these 'Critical' models are governed. The goal is to ensure that the tools used to secure our systems do not become the primary weapons used to break them.