OpenAI Launches GPT-6 Cyber With 'Critical' Cybersecurity Rating and Restricted Access

OpenAI has released GPT-6 Cyber, a specialized model capable of finding software vulnerabilities and developing working exploits. Access is restricted to verified security researchers amid concerns over AI-enabled cyberattacks.

[For more news, click here]

Security researchers are increasingly turning to AI to examine complex software, identify vulnerabilities, and test whether weaknesses can be exploited. But the ability to develop working exploits creates a challenge for AI developers: tools that help defenders could also make sophisticated attacks easier to carry out.

OpenAI has released GPT-6 Cyber, a specialized model for vulnerability research and authorized penetration testing, initially making it available to a small group of verified customers.

The company classifies the model as "Critical" in cybersecurity under its Preparedness Framework, reflecting its advanced capabilities rather than an assessment of the model's own security.

How Powerful Is GPT-6 Cyber?

According to OpenAI's official announcement, GPT-6 Cyber can identify software vulnerabilities, reproduce security flaws, and develop working exploits to help researchers determine their impact before patches are deployed.

On ExploitGym, a benchmark testing whether models can exploit specified vulnerabilities, GPT-6 Cyber scored 48.5%, compared with 32.1% for GPT-5.6 Cyber.

In a separate controlled evaluation involving known vulnerabilities in applications including GitLab and PostgreSQL, the new model completed 95% of remote exploitation tasks.

However, performance varies across evaluations.

External testing by cybersecurity evaluation firm Irregular found that GPT-6 Cyber solved 90 of 288 FrontierCyber challenges, slightly fewer than the 96 solved by GPT-6 Astra. Neither model completed the benchmark's 15 most difficult challenges.

These results suggest substantial improvements in certain cybersecurity tasks, but they do not establish that the model can compromise every well-protected system.

Firefox and Linux Findings Demonstrate Real-World Capabilities

OpenAI says its researchers used GPT-6 Cyber to uncover 81 previously unknown vulnerabilities in Firefox that Mozilla classified as high severity.

The findings were disclosed to Mozilla, which has begun releasing fixes. The model also helped develop proposed patches for security researchers to review.

Brian Grinstead, a distinguished Firefox engineer, said the work "helped us ship fixes to millions of users."

In separate Linux kernel research, OpenAI reported that the model reproduced 1,724 vulnerabilities and developed working exploits for 948 of them.

The company also disclosed 304 critical findings to Linux maintainers, alongside candidate patches.

These figures demonstrate the model's ability to support vulnerability investigation at scale. However, the 948 working exploits should not be interpreted as 948 newly discovered zero-day vulnerabilities.

The findings also illustrate the distinction between defensive and offensive cybersecurity capabilities.

Developing an exploit can help authorized researchers verify a vulnerability and understand its consequences. The same technique, used without authorization, can enable malicious activity.

OpenAI Limits Access to Verified Security Researchers

Given these risks, GPT-6 Cyber is initially available through Daybreak Red, OpenAI's program for advanced, authorized security testing.

Selected customers can access the model through the API and Codex, with availability expected to expand to other approved Daybreak Red users.

According to the model's technical system card, the access requirements include identity verification, multifactor authentication, managed devices, and restrictions on the networks permitted to use the model.

OpenAI has also introduced monitoring systems designed to identify potentially harmful activity and actions outside a researcher's authorized scope.

In one internal evaluation, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of tests without deployment safety prompts. With those prompts, the model stayed within the permitted scope in all observed tests.

However, these results come from controlled evaluations and do not guarantee equivalent behavior in real-world environments.

Additional protections focus on critical infrastructure, including industrial control systems and essential services, where unauthorized activity could result in physical damage or service disruptions.

What This Means for Cybersecurity

The release illustrates how AI cybersecurity tools are evolving from systems that identify potential weaknesses into agents capable of validating vulnerabilities through sustained technical work.

For security teams, this could shorten the time required to investigate complex flaws and develop appropriate fixes.

However, increasingly capable models also raise questions about access controls, independent safety testing, and the possibility that similar capabilities could become available to malicious actors.

The broader challenge is ensuring that improvements in automated vulnerability research strengthen defenses faster than they expand opportunities for attackers.

Ultimately, the value of advanced cybersecurity AI will depend not only on its ability to discover vulnerabilities, but also on how effectively developers control its use and how quickly organizations can act on its findings.