OpenAI Announces GPT-6 Cyber for Advanced Cybersecurity Research

OpenAI announced GPT-6 Cyber on November 5, 2026, a specialized cybersecurity model designed to identify software vulnerabilities, verify whether they can be exploited, and develop fixes. The company reported significant results from security testing involving Firefox and the Linux kernel, including the discovery of 81 new high-severity vulnerabilities in Firefox and the development of 948 working exploits for Linux vulnerabilities.

Introducing GPT‑6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment.

The model powered the Linux kernel research we announced yesterday, and security teams at companies like Mozilla, Palo Alto Networks, Calif, and… pic.twitter.com/34uAeNnhwP

– OpenAI (@OpenAI) November 5, 2026

Built on GPT-6 Astra, the new model received additional training focused on cybersecurity tasks and reducing unnecessary refusals of legitimate defensive requests. It combines vulnerability research, penetration testing, and exploit development capabilities, while operating under controls designed to keep its activities within the scope authorized by security researchers.

OpenAI has begun rolling out GPT-6 Cyber through its Daybreak Red program to a group of approved organizations, with access expanding gradually to other eligible customers.

What Is OpenAI GPT-6 Cyber?

GPT-6 Cyber is a specialized version of GPT-6 Astra developed by OpenAI for advanced cybersecurity work. Its capabilities include analyzing source code, identifying security weaknesses, reproducing vulnerabilities, testing whether they can be exploited, and preparing fixes within authorized environments.

Unlike general-purpose coding assistants, GPT-6 Cyber is designed to handle a broader security research workflow. Rather than simply reviewing code and pointing out a potential flaw, it can investigate the issue, test different approaches, build a functional exploit to demonstrate the vulnerability’s impact, and develop a software patch to address the underlying problem.

OpenAI used additional reinforcement learning to train the model on tasks such as reproducing software crashes, searching large codebases for vulnerabilities, and developing executable exploits. The training also focused on reducing unnecessary refusals of specialized technical requests while maintaining restrictions against unauthorized activities.

GPT-6 Cyber Finds 81 Vulnerabilities in Firefox

One of the most notable results came from a security research campaign targeting Mozilla Firefox, one of the world’s most widely used open-source browsers and a project that undergoes extensive security testing.

The campaign resulted in GPT-6 Cyber identifying 81 previously undiscovered vulnerabilities that Mozilla classified as high severity. The findings exceeded the results of earlier security assessments conducted on the browser using multiple generations of advanced AI models.

The work went beyond identifying potential weaknesses. GPT-6 Cyber prepared reproducible evidence and technical tests that allowed Mozilla engineers to verify the reported issues directly and evaluate proposed fixes. Mozilla subsequently began releasing patches for a number of the vulnerabilities identified through the campaign.

The research also uncovered 30 previously known security issues whose actual severity was higher than their earlier assessments suggested. This gave developers a stronger basis for reassessing remediation priorities and addressing critical weaknesses across the widely used browser.

948 Working Exploits for Linux Kernel Vulnerabilities

OpenAI also tested GPT-6 Cyber against the Linux kernel, the core component that powers Linux distributions and a substantial part of the world’s server infrastructure.

During the campaign, the model reproduced 1,724 security vulnerabilities and developed working exploits for 948 of them, with the goal of escalating privileges within the system.

These results demonstrate the model’s ability to move beyond theoretical vulnerability detection and produce functional exploits. This process involves understanding the conditions behind a security flaw, reproducing the issue, and testing its practical impact in an operating environment.

The campaign produced 304 critical findings that were reported to Linux maintainers alongside proposed software fixes. Work on addressing the remaining vulnerabilities continued in coordination with kernel developers.

The comparison with the previous GPT-5.6 Sol model also highlights the scale of the improvement. GPT-6 Cyber developed more than five times as many exploits in the same testing environment, while its predecessor produced only 166.

Benchmark Performance and Security Metrics

OpenAI reported substantial improvements across several benchmarks designed to measure vulnerability discovery and exploit development capabilities.

• ExploitGym: This benchmark measures a model’s ability to exploit a specific vulnerability in a remote target. GPT-6 Cyber achieved a success rate of 48.5%, compared with 32.1% for GPT-5.6 Cyber, representing an improvement of approximately 51%.
• Web application security testing: In a black-box vulnerability discovery test, where the model does not have access to the application’s source code, GPT-6 Cyber achieved a recall rate of 40.3%, compared with 32% for GPT-6 Astra.
• RemoteExploitBench: This benchmark simulates production environments involving platforms such as GitLab, PostgreSQL, and Jenkins. GPT-6 Cyber completed approximately 95% of the tasks. By comparison, GPT-6 Astra operating with Daybreak Blue settings refused all the tasks under its existing security restrictions.

Together, these results show how far AI agents have progressed in carrying out practical security testing and exploitation tasks.

Why Did OpenAI Classify GPT-6 Cyber as Critical?

OpenAI classified GPT-6 Cyber at the Critical level for cybersecurity capabilities under its Preparedness Framework.

The classification reflects the model’s advanced offensive security capabilities, particularly its ability to produce functional exploits. Its predecessor, GPT-5.6 Cyber, was classified at the High level, while GPT-6 Astra also carries the Critical classification.

This level of capability comes with strict deployment and verification requirements. Organizations using a model capable of developing working exploits need reliable systems for verifying operator identities, defining authorized targets, and monitoring security operations.

How OpenAI Controls GPT-6 Cyber’s Security Risks

OpenAI introduced technical safeguards designed to keep GPT-6 Cyber within the boundaries of an assigned security task and prevent it from targeting neighboring systems that have not been authorized for testing.

One of the tests used to evaluate these safeguards was ExploitGym Honeypot. The benchmark places exposed, deliberately vulnerable services inside a testing environment to measure whether a model attempts to access systems outside its authorized scope.

GPT-6 Cyber avoided targeting out-of-scope services in 99.93% of cases without additional safety instructions. That figure increased to 100% when the additional security guidance was enabled.

The deployment controls also include approved IP address lists, API key protection, and enterprise security requirements such as single sign-on (SSO), multi-factor authentication (MFA), and managed devices.

OpenAI also partnered with Dragos, a company specializing in industrial cybersecurity, to test the model in an environment simulating an electrical facility. The assessment examined its ability to interpret industrial network components and identify threats that could affect control systems and sensitive equipment.

Deployment and Access via Daybreak Red

OpenAI is making GPT-6 Cyber available through its API and Codex to a selected group of approved Daybreak Red customers. The program is designed for advanced security teams that need to conduct penetration tests and develop working exploits in authorized environments.

Daybreak Red provides broader operational capabilities than Daybreak Blue. While Blue focuses on defensive security assessments and patch validation, Red supports advanced penetration testing and the development of complete attack scenarios.

OpenAI has also made GPT-5.6 Cyber available to all Daybreak Red participants, while gradually expanding access to GPT-6 Cyber. Organizations must meet security screening and identity verification requirements before receiving access.

For now, GPT-6 Cyber remains restricted to the program rather than being offered as a publicly available product with general pricing.

What GPT-6 Cyber Means for the Future of Cybersecurity

The launch of GPT-6 Cyber marks a significant development in AI-powered cybersecurity. AI is moving beyond the role of an advisory assistant toward autonomous agents capable of handling major parts of the security research process, from finding vulnerabilities and testing exploits to preparing software patches.

The results from Firefox and Linux demonstrate the growing practical role of AI in vulnerability research. By automating complex stages of security testing, models like GPT-6 Cyber can help security teams examine large codebases, identify weaknesses faster, and prioritize fixes before attackers exploit them.

For software developers and organizations responsible for critical infrastructure, this creates a new approach to security testing: AI agents can perform extensive technical investigations while human security teams define the scope, verify findings, and oversee remediation.