OpenAI has introduced GPT-6 Cyber, a specialised artificial intelligence model designed to help cybersecurity professionals identify software vulnerabilities, test potential exploits and develop fixes. The release marks a significant development in AI-assisted security research, but it also highlights growing concerns about the risks associated with increasingly capable automated hacking tools.
The company announced the model on November 5, 2026, describing it as a major advancement over its previous cybersecurity-focused model, GPT-5.6 Cyber. Reuters also reported the launch on November 5, confirming that initial access would be limited to a small group of vetted customers. The announcement has since attracted attention across the technology and cybersecurity industries.
OpenAI has classified GPT-6 Cyber as Critical in cybersecurity under its Preparedness Framework. This is the company’s internal risk classification, indicating that the model has reached a level of capability that requires additional safeguards and restricted access. It should not be interpreted as an independent security certification or proof that the model can successfully compromise every type of protected system.
According to OpenAI’s published evaluations, GPT-6 Cyber achieved a score of 48.5% on ExploitGym, a benchmark designed to measure a model’s ability to exploit specified vulnerabilities in remote targets. Its predecessor, GPT-5.6 Cyber, scored 32.1% on the same evaluation. OpenAI also reported that the new model achieved 40.3% vulnerability recall in a separate web penetration-testing evaluation, compared with 32% for GPT-6 Astra.
The company says the model has already contributed to security research involving widely used software. OpenAI reports that GPT-6 Cyber helped identify 81 previously unknown high-severity vulnerabilities in Firefox and develop working exploits for 948 Linux kernel vulnerabilities. It also says candidate patches and findings were shared with relevant maintainers, with Mozilla having released fixes for many of the reported Firefox vulnerabilities.
These figures are significant, but they require context. They are results published by OpenAI, rather than a comprehensive independent assessment of the model’s performance across all real-world environments. OpenAI’s system card does include evaluations conducted by external organisation Irregular. Those results show a more mixed picture: GPT-6 Cyber solved 90 of 288 challenges in one evaluation, compared with 96 for GPT-6 Astra, and neither model successfully completed any of the 15 highest-difficulty challenges. The external testing therefore provides useful evidence of capability while also showing that performance varies considerably depending on the task and testing conditions.
A central feature of the release is OpenAI’s attempt to address the security risks created by more capable AI systems. A model that can identify weaknesses and develop working exploits can help defenders find problems before attackers do. However, the same technical capabilities could be misused to compromise networks, steal information or disrupt essential services.
OpenAI says GPT-6 Cyber has received specialised training to follow the boundaries set by authorised security researchers. In its ExploitGym Honeypot evaluation, the company reports that the model avoided exploiting systems outside the assigned scope in 99.93% of tests without deployment safety prompts, rising to 100% when those prompts were enabled. These are controlled evaluation results, not a guarantee that the model will never take an unauthorised action in a live environment.
The company has also introduced additional safeguards intended to restrict dangerous activity, particularly actions involving critical infrastructure such as power plants, water treatment facilities and industrial control systems. Its system card describes access requirements that include identity verification and security controls for approved users and organisations. The model is initially being made available through OpenAI’s Daybreak Red programme to selected, approved customers using the API and Codex, rather than being released as a general-purpose tool for unrestricted public use.
The launch has implications well beyond the cybersecurity industry. For businesses, AI-assisted vulnerability research could make it easier to identify weaknesses in applications, cloud infrastructure and internal systems. Organisations with limited security budgets may eventually benefit from tools that help security teams examine larger codebases and investigate potential threats more efficiently. However, adopting these systems will still require qualified professionals, controlled testing environments and clear rules governing access to company data and infrastructure.
For software developers and open-source maintainers, the technology could accelerate security reviews and help uncover vulnerabilities in projects that have undergone years of testing. At the same time, the growing availability of automated exploit-development capabilities could shorten the time between the discovery of a vulnerability and attempts to abuse it. This makes timely software updates, responsible vulnerability disclosure and effective patch management increasingly important.
Governments and critical infrastructure operators face another challenge: ensuring that increasingly capable AI systems are used to strengthen national cybersecurity rather than create new opportunities for disruption. Restrictions on access may reduce exposure, but they cannot eliminate the broader risks associated with the spread of advanced cyber capabilities.
The release also raises questions for investors and technology companies. Cybersecurity is becoming an increasingly important area of competition among AI developers, software providers and security firms. Companies that successfully integrate AI into vulnerability detection, incident response and secure software development could gain an advantage. However, the commercial benefits will depend on whether these systems can deliver reliable results, maintain appropriate safeguards and earn the confidence of enterprise customers.
OpenAI’s announcement is an important step in the development of AI-powered cybersecurity, but the technology’s long-term value will depend on more than benchmark scores. Its performance in real-world deployments, the effectiveness of its restrictions and the ability of independent researchers to evaluate its claims will all matter.
As AI systems become more capable of identifying and exploiting software weaknesses, cybersecurity teams will need to adapt their defences accordingly. GPT-6 Cyber illustrates the opportunity to use advanced AI to find and fix vulnerabilities faster, while demonstrating why access controls, human oversight and responsible deployment are becoming essential parts of the technology.
