Artificial intelligence is becoming a more active part of cybersecurity research, helping specialists examine complex software and identify weaknesses that may be difficult to detect through conventional testing. OpenAI’s latest release, OpenAI GPT-6 Cyber, is designed specifically for this work.
Announced on November 5, 2026, the model focuses on vulnerability discovery, exploit development, and security testing within boundaries established by authorized researchers. Unlike a general-purpose chatbot, OpenAI GPT-6 Cyber is built for advanced security workflows, including vulnerability discovery, exploit development, penetration testing, and patch validation.
Its initial research results involve two widely used software environments, Mozilla Firefox and the Linux kernel, demonstrating how AI-assisted security research can help developers identify, reproduce, and address complex software weaknesses.
GPT-6 Cyber Identifies 81 Firefox Vulnerabilities
One of the most notable results came from a vulnerability research campaign involving Mozilla Firefox. OpenAI reports that GPT-6 Cyber identified 81 previously unknown vulnerabilities, supported by reproducible tests and evidence. Mozilla subsequently classified all 81 findings as high severity.
The research demonstrates the model’s ability to examine extensively tested software codebases and produce findings that developers can independently reproduce and investigate. In addition to the 81 newly identified vulnerabilities, GPT-6 Cyber also found 30 previously known Firefox issues that were more severe than initially reported.
OpenAI disclosed the findings to Mozilla and provided proposed patches to help its engineers develop fixes. Mozilla began landing fixes shortly after receiving the reports and has already released updates addressing many of the reported vulnerabilities.
However, this does not mean that all 81 vulnerabilities have been patched. The findings show how AI-assisted research can help developers reassess existing security issues and discover previously unknown weaknesses.
Linux Research Produced 948 Working Exploits
OpenAI also used GPT-6 Cyber to investigate vulnerabilities in the Linux kernel.
During the research campaign, the model reproduced 1,724 vulnerabilities and developed 948 working privilege-escalation exploits. These exploits demonstrated how an attacker could potentially gain higher system privileges by exploiting weaknesses in the kernel.
The figures represent different stages of the research: reproducing a vulnerability does not necessarily mean that a working exploit has been developed.
The results also show a substantial improvement over the previous generation. Using the same research harness, GPT-5.6 Sol successfully exploited 166 Linux kernel vulnerabilities, compared with 948 for GPT-6 Cyber. This represents more than a fivefold increase in successful exploit development under the reported testing conditions.
These figures describe different stages of the research: reproducing a vulnerability and successfully developing an exploit are not the same result. OpenAI separately disclosed 304 critical findings to Linux maintainers, together with candidate patches to support remediation.
These disclosures represent a separate category of research outcomes from the 1,724 reproduced vulnerabilities and 948 successfully exploited vulnerabilities. We are still addressing other reported issues. Importantly, the 948 figure refers to working privilege-escalation exploits developed during the research campaign, not 948 newly discovered zero-day vulnerabilities. Instead, it represents the number of working exploits developed during the research campaign.
These results show the model’s ability to move beyond identifying suspicious code and test whether vulnerabilities can be exploited in controlled research environments. That capability can help defenders understand the practical impact of software weaknesses and prioritize fixes.
ExploitGym Shows Improvement Over GPT-5.6 Cyber
OpenAI reported a score of 48.5% for GPT-6 Cyber on ExploitGym, a benchmark that evaluates a model’s ability to exploit a specified vulnerability in a remote target.
GPT-5.6 Cyber scored 32.1% in the same comparison. This represents an improvement of 16.4 percentage points, equivalent to approximately 51% in relative performance. The result highlights the model’s improved ability to turn vulnerability knowledge into working exploits under benchmark conditions. However, this result comes from a specific benchmark, not a measure of general cybersecurity accuracy or guaranteed success against real-world systems.
OpenAI also evaluated GPT-6 Cyber on a web penetration-testing benchmark in which models examined web applications without access to their underlying source code. GPT-6 Cyber achieved a vulnerability-recall score of 40.3%, compared with 32.0% for GPT-6 Astra. The evaluation measures the model’s ability to identify vulnerabilities through external testing rather than direct examination of application source code.
OpenAI also introduced RemoteExploitBench, a new evaluation designed to measure a model’s ability to exploit known vulnerabilities in remote applications running in controlled environments. GPT-6 Cyber completed 95% of the evaluated tasks, while GPT-6 Astra refused all the tasks.
The benchmark uses known vulnerabilities in controlled testing environments, rather than measuring success against live production systems. The difference in results also reflects the model’s refusal behavior, as GPT-6 Astra declined the tasks instead of attempting them. Therefore, the 95% completion rate should be understood as a benchmark result, not a guaranteed success rate against real-world targets.
Why OpenAI Classifies GPT-6 Cyber as Critical
Under its Preparedness Framework, OpenAI has classified GPT-6 Cyber as Critical in cybersecurity, compared with the High classification assigned to GPT-5.6 Cyber.
These are OpenAI’s internal capability-risk classifications, reflecting the model’s potential to perform advanced cybersecurity tasks involving vulnerability discovery, exploit development, and penetration testing. They should not be interpreted as independent security certifications or assessments of the model’s overall reliability.
Greater capability also creates risks. A model that helps defenders find weaknesses could potentially assist malicious actors if its access is not controlled. OpenAI says it has introduced additional cybersecurity safeguards to reduce misuse and keep the model’s actions within the permissions researchers grant. These include
Cyber-specific monitoring:
Cyber-specific monitoring to identify activity outside authorized testing boundaries, protections against compromised or stolen credentials, and additional checks against high-risk assistance involving critical infrastructure.
Credential protection:
Includes monitoring for compromised credentials and restrictions on API access.
IP allowlisting:
Daybreak Red API access requires IP allowlisting by default, with limited exceptions, to restrict requests to approved networks.
Critical infrastructure safeguards:
Additional checks address high-risk assistance involving essential systems and industrial controls.
Sandboxing and human oversight:
OpenAI also recommends isolated testing environments, sandboxing, action reviews, and human oversight for high-risk security workflows.
OpenAI assessed the model’s ability to respect authorized testing boundaries through its ExploitGym Honeypot evaluation. Without deployment safety prompts, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of tested cases, compared with 94.49% for GPT-5.6 Cyber.
When deployment safety instructions were enabled, GPT-6 Cyber stayed within the authorized scope in all tested cases. These results indicate improved alignment with researchers’ permissions, although they apply only to the specific evaluation conditions and do not guarantee identical behavior in every real-world deployment.
Daybreak Blue and Daybreak Red Explain Access Levels
OpenAI’s Daybreak program provides different levels of cybersecurity access through its Blue and Red tiers. Daybreak Blue is designed for verified security defenders who need to identify and validate vulnerabilities, test potential fixes, and strengthen software security.
Daybreak Red provides access to more advanced capabilities, including exploit development and authorized security testing against remote targets.
GPT-6 Cyber is being offered through the more restricted Daybreak Red program because of its advanced vulnerability discovery and exploit-development capabilities.
OpenAI has also introduced Private Safety Processing (PSP) for eligible sensitive API workflows. The system performs automated safety checks to identify potential misuse. For API customers using Zero Data Retention, PSP is generally required for eligible workflows and reviews encrypted records retained in customer storage for 30 days.
OpenAI says this process does not expose prompts and responses to its personnel. The system is intended to support automated safety monitoring while accommodating stricter data-retention requirements. OpenAI has also introduced Daybreak Red for Critical Infrastructure, a dedicated access group for approved defenders working with essential systems.
The program supports authorized security testing of critical infrastructure, including power plants, water treatment facilities, operational technology, and industrial control systems. Approved participants receive permission to test these environments while remaining subject to safeguards designed to prevent unauthorized or potentially dangerous actions.
Private Safety Processing Adds Another Security Layer
OpenAI GPT-6 Cyber is initially available through the API and Codex only to a selected group of approved Daybreak Red customers. Access is not automatically available to all API users, Codex customers, or ChatGPT subscribers.
OpenAI plans to expand availability to all approved Daybreak Red customers progressively, although it has not announced a specific public release date for general access. The model is already being used by security teams at several technology companies.
OpenAI identifies Cisco, Calif and CodeRabbit among the organizations using OpenAI GPT-6 Cyber for vulnerability research and code review. These early deployments suggest how specialized AI models could help security professionals examine complex codebases, validate potential weaknesses, and improve software security reviews.
OpenAI has also used GPT-6 Cyber alongside other models in controlled security research to identify and address weaknesses in its own infrastructure. The company reports that researchers used AI agents to investigate vulnerabilities, including weaknesses that could allow attackers to escape microvirtual machines and execute code on underlying hosts.
OpenAI subsequently used the findings to patch software, strengthen network restrictions, and report issues to affected maintainers and vendors. These results came from research involving multiple OpenAI models and should not be attributed exclusively to OpenAI GPT-6 Cyber.
The restricted program is intended for trusted security professionals who need advanced capabilities for authorized testing and vulnerability research.
OpenAI’s safeguards include identity verification, organizational security requirements, API IP allowlisting, and monitoring for activity outside approved permissions. The company also recommends sandboxing and human oversight for high-risk security workflows.
What GPT-6 Cyber Could Mean for Software Security
The model’s research results point to a growing role for AI in vulnerability discovery and software defense.
Security teams could use tools like GPT-6 Cyber to examine large codebases, validate potential exploits, and help develop fixes more efficiently. This may be particularly useful for open-source projects and organizations that need to investigate complex security issues. However, stronger exploit-development capabilities also make access controls and responsible disclosure more important.
The value of such systems will depend not only on how many vulnerabilities they can identify, but also on whether their findings can be verified, reported and fixed before attackers take advantage of them. Despite these advances, benchmark performance and controlled research results should not be treated as guarantees of success against real-world systems.
Reproducing a vulnerability or developing a working exploit in an isolated environment does not prove the same technique will work against a live, fully protected target. Similarly, vulnerability discovery results depend on the software examined, testing conditions, and verification process. Human review, responsible disclosure and timely patch deployment remain essential to translating AI-assisted research into practical security improvements.
The Bottom Line
OpenAI GPT-6 Cyber marks a significant development in specialized AI-assisted security research. Its Firefox and Linux results show its potential to identify vulnerabilities and develop working exploits at scale, while its Critical cybersecurity classification highlights the risks of those capabilities.
For now, restricted access and additional safeguards position the model as a tool for approved security researchers rather than a general-purpose chatbot.