OpenAI has unveiled GPT-6 Cyber, a specialized artificial intelligence model designed to identify software vulnerabilities, develop working exploits, and help cybersecurity teams strengthen their defenses before attackers can take advantage of weaknesses.
Announced on November 5, 2026, the model represents a significant advance in AI-powered security research. It goes beyond detecting suspicious code, allowing researchers to investigate vulnerabilities, demonstrate their potential impact, and develop possible fixes.
The numbers are striking. OpenAI researchers used GPT-6 Cyber to uncover 81 previously unknown high-severity vulnerabilities in Mozilla Firefox and develop working exploits for 948 Linux kernel vulnerabilities.
OpenAI has classified the model as having Critical cybersecurity capabilities under its Preparedness Framework, its highest cybersecurity capability designation.
The company is also taking a cautious approach to access. Rather than releasing GPT-6 Cyber to everyone, OpenAI is initially limiting availability to approved security researchers and organizations through its Daybreak Red program.
GPT-6 Cyber Delivers a Major Leap in AI-Powered Cybersecurity
GPT-6 Cyber builds on the capabilities of GPT-5.6 Cyber, with specialized training focused on vulnerability discovery, exploit development, and authorized penetration testing. The model can examine complex software systems, investigate potential weaknesses, and determine whether those weaknesses could lead to successful attacks.
OpenAI reported a substantial performance improvement on ExploitGym, a cybersecurity benchmark that evaluates whether AI models can exploit known vulnerabilities in controlled environments. GPT-6 Cyber achieved a success rate of 48.5%, compared with 32.1% for GPT-5.6 Cyber. That translates to approximately 51% better relative performance.
The model also demonstrated stronger capabilities in web penetration testing. During evaluations involving applications where the AI could not access the underlying source code, GPT-6 Cyber identified 40.3% of vulnerabilities, outperforming GPT-6 Astra’s 32% recall rate.
These improvements matter because identifying a vulnerability does not necessarily mean a security researcher understands how it could be exploited. GPT-6 Cyber can investigate that connection, helping defenders distinguish serious security weaknesses from issues that may pose less immediate danger.
OpenAI Discovers 81 Previously Unknown Firefox Vulnerabilities
Mozilla Firefox became one of the clearest demonstrations of GPT-6 Cyber’s capabilities. Despite extensive security testing and previous analysis using advanced AI models, OpenAI researchers identified 81 previously unknown vulnerabilities that Mozilla classified as high severity.
The model also uncovered evidence suggesting that 30 previously identified security issues were more severe than initially understood. OpenAI shared its findings with Mozilla, providing reproducible tests and proposed patches to help engineers investigate and address the problems.
Mozilla has already released fixes for many of the vulnerabilities identified during the research.
The findings are particularly notable because Firefox is a mature browser with a long history of security reviews. Discovering dozens of serious weaknesses in such a heavily examined codebase suggests that advanced AI models may be capable of finding problems that conventional security processes have missed.
For software developers, this could change how vulnerability assessments are conducted. Instead of relying entirely on manual code reviews and traditional automated scanners, security teams may increasingly use AI agents to investigate complicated software behavior and recommend fixes.
GPT-6 Cyber Develops Working Exploits for 948 Linux Kernel Vulnerabilities
Linux was another major testing ground for OpenAI’s cybersecurity model. Researchers used GPT-6 Cyber to reproduce 1,724 Linux kernel vulnerabilities and successfully develop working exploits for 948 of them.
The improvement over earlier models was substantial. Using the same research framework, GPT-5.6 Sol developed working exploits for 166 vulnerabilities, meaning GPT-6 Cyber achieved more than five times that result.
OpenAI also disclosed 304 critical findings to Linux maintainers, providing candidate patches to support remediation efforts.
The significance extends well beyond the Linux development community. Linux powers much of the infrastructure behind cloud computing, enterprise servers, networking equipment, and countless connected devices.
Security weaknesses in the kernel can therefore affect systems used by businesses, governments, and technology providers worldwide.
What makes GPT-6 Cyber particularly interesting is its ability to connect vulnerability discovery with practical verification. Rather than simply identifying potentially unsafe code, the model can help researchers determine whether a weakness creates a genuine security risk.
That could make vulnerability management faster and more precise, although organizations will still need experienced security professionals to validate findings and implement appropriate fixes.
OpenAI Strengthens Safety Controls to Prevent Unauthorized Attacks
Powerful cybersecurity tools come with an obvious concern. The same techniques used to uncover software vulnerabilities can potentially support malicious activity when placed in the wrong hands.
OpenAI says GPT-6 Cyber received specialized training to follow the boundaries established by authorized researchers. The model is designed to distinguish between permitted security testing and actions that fall outside an approved assignment.
During an evaluation called ExploitGym Honeypot, researchers tested whether the model would exploit additional vulnerable services that were deliberately placed outside its authorized scope.
Without additional deployment safety instructions, GPT-6 Cyber avoided exploiting those services in 99.93% of evaluated cases. GPT-5.6 Cyber achieved 94.49% under the same conditions.
With deployment safety instructions enabled, GPT-6 Cyber remained within its assigned boundaries in 100% of the evaluated cases.
These results suggest meaningful improvements in the model’s ability to respect testing restrictions. Still, OpenAI acknowledges that stronger safeguards remain necessary as AI cybersecurity capabilities continue advancing.
The company has introduced additional monitoring, account security requirements, and restrictions designed to reduce the risk of unauthorized activity.
Critical Infrastructure Security Becomes a Major Focus
The potential impact of AI-assisted cybersecurity becomes even more serious when industrial systems enter the picture. Power stations, water treatment facilities, manufacturing plants, and other essential services increasingly depend on connected software and operational technology.
OpenAI partnered with industrial cybersecurity company Dragos to evaluate how advanced AI models perform in controlled industrial environments.
During testing involving a simulated power plant, researchers observed that an AI model could distinguish operational technology assets from ordinary IT systems and identify two previously unknown vulnerabilities in engineering software associated with electrical protection equipment.
Those weaknesses involved access to protection settings and breaker control logic, areas where unauthorized interference could have serious consequences in real-world environments.
The findings demonstrate why advanced cybersecurity AI requires careful deployment. Such models could help infrastructure operators identify dangerous weaknesses before attackers discover them, but they could also lower the technical barriers to researching industrial systems.
OpenAI has responded by strengthening safeguards against assistance that could enable serious physical damage, disruption of essential services, or other dangerous outcomes.
The company is also expanding support for approved critical infrastructure defenders through Daybreak Red for Critical Infrastructure and its broader Daybreak initiatives.
Cybersecurity Companies Begin Using GPT-6 Cyber for Security Research
GPT-6 Cyber is already attracting attention from established cybersecurity companies and software development teams.
OpenAI says organizations including Cisco, Palo Alto Networks, SentinelOne, Calif, and CodeRabbit have evaluated or used the model to investigate vulnerabilities, review software, and strengthen security operations.
CodeRabbit reported that GPT-6 Cyber identified 66% more high- and medium-severity security issues than its existing production model during internal benchmarking.
The results suggest that specialized AI systems could become valuable tools for developers working with increasingly complex applications.
Modern software frequently relies on interconnected libraries, external services, and large codebases. Security weaknesses may involve several components interacting in unexpected ways, making them difficult to discover through conventional testing alone.
GPT-6 Cyber is designed to investigate these relationships over longer research sessions, helping security teams identify vulnerabilities that require multiple steps to reproduce.
OpenAI has also used the model to strengthen its own infrastructure. During controlled security testing, AI agents helped researchers identify weaknesses that could allow attackers to move beyond isolated computing environments.
Those findings supported additional patches, network restrictions, and disclosures to affected software maintainers.
GPT-6 Cyber Access Is Restricted Through OpenAI Daybreak Red
OpenAI is initially making GPT-6 Cyber available through its API and Codex to a selected group of approved Daybreak Red customers.
The program is intended for authorized cybersecurity activities, including advanced penetration testing, vulnerability research, and exploit development.
Access comes with security requirements designed to reduce misuse. Organizations must meet identity verification and account protection standards, while individual researchers face additional eligibility and authentication requirements.
OpenAI also requires IP allowlisting for Daybreak Red API access by default, limiting requests to approved network addresses unless an exception applies.
The company plans to expand GPT-6 Cyber availability to additional approved Daybreak Red customers. Meanwhile, GPT-5.6 Cyber is being made available to all Daybreak Red customers, and Daybreak Blue access is expanding for eligible defensive security researchers.
This controlled rollout reflects the unusual position OpenAI now faces. It wants security teams to benefit from increasingly capable AI systems, but unrestricted access could also create opportunities for attackers.
GPT-6 Cyber Could Change How Organizations Find and Fix Vulnerabilities
The arrival of GPT-6 Cyber signals a shift in how artificial intelligence may influence cybersecurity operations.
Traditional security tools have long helped developers scan code and detect suspicious behavior. Advanced AI models are beginning to take on more complicated research tasks, including investigating vulnerabilities, testing exploitation scenarios, and preparing possible fixes.
For organizations maintaining large software platforms, these capabilities could reduce the time required to identify and understand serious security weaknesses.
There is another side to the development, however.
As AI models become more capable of finding exploitable vulnerabilities, attackers may also gain access to increasingly sophisticated methods for identifying weaknesses in widely used software.
That creates pressure on developers, infrastructure operators, and cybersecurity teams to improve how quickly they investigate reports and deploy security updates.
OpenAI’s research involving Firefox and Linux shows how much progress AI-assisted vulnerability discovery has already made.
Whether that progress ultimately benefits defenders more than attackers will depend on access controls, responsible disclosure, effective monitoring, and the speed at which organizations respond to newly discovered threats.
GPT-6 Cyber may be one of the clearest signs yet that artificial intelligence is moving beyond assisting cybersecurity professionals and becoming an active participant in complex security research.
The challenge now is ensuring that its growing capabilities strengthen digital defenses rather than create new risks.
Sources
OpenAI — GPT-6 Cyber: Advancing the Frontier of Aligned Cybersecurity Agents
https://openai.com/index/introducing-gpt-6-cyber/
OpenAI — GPT-6 Cyber System Card
https://deploymentsafety.openai.com/gpt-6-cyber
Reuters — OpenAI Launches GPT-6 Cyber With Critical Cybersecurity Rating
https://www.investing.com/news/stock-market-news/openai-launches-gpt6-cyber-with-critical-cybersecurity-rating-5005229
Axios — OpenAI Unveils Cyber Model With Unprecedented Hacking Skills
https://www.axios.com/2026/11/05/openai-linux-cyber-model
