OpenAI on November 5, 2026, introduced a new artificial intelligence model, GPT-6 Cyber. It is a specially designed version that is trained to detect exploitable software vulnerabilities, validate them with working attacks, and adhere to the limits set by security researchers. OpenAI assesses the model’s risk level using its internal Preparedness Framework, in which GPT-6 Cyber received the highest rating of „Critical.“ The previous version, GPT-5.6 Cyber, was rated „High,“ so this jump represents a significant increase in power and risk.
Critical appraisal and its significance
„The Preparedness Framework is OpenAI’s internal system for classifying a model’s potential vulnerabilities. “Critical„ is the highest risk level, as noted by British technology portal The Register. This means that the model has the potential to cause serious security threats if misused. For this reason, GPT‑6 Cyber is not included in the ChatGPT service, and access to it is restricted.
Access via the Daybreak Red app
GPT-6 Cyber is currently available only through the OpenAI API and Codex, and only to a limited number of customers approved by the Daybreak Red program. This program is designed for offensive security tasks, such as authorized penetration testing. OpenAI plans to expand access to all Daybreak Red customers in the future. The model uses IP whitelists by default, constantly monitors for potential credential theft, and blocks assistance related to attacks on energy, water treatment, or industrial control systems.
Test results and practical application
Experiments with ExploitGym, a platform where a model must exploit a specific software vulnerability on a remote server, showed that GPT-6 Cyber completed 48.5 % tasks, compared to 32.1 % for its predecessor. This represents a significant performance jump. In addition, OpenAI claims that the model did not cause external service exploitation in 99.93 % cases when a special „deployment“ command was not expressed. Details of the remaining 0.07 % cases were not disclosed.
Specific security findings
OpenAI researchers used GPT-6 Cyber to identify 81 high-risk vulnerabilities in the Firefox browser, 41 of which were included in the Firefox 159 security advisory released on October 27, 2026. Mozilla credited OpenAI with 167 of the vulnerabilities fixed. Firefox is currently at version 159.0.1, with older LTS versions 153.6 and 115.44.
In the Linux kernel, GPT‑6 Cyber recovered 1,724 vulnerabilities and created exploits in 948 cases, giving the attacker root privileges. OpenAI reported 304 critical findings to the kernel maintainers along with blacklist patches, and further patching work is ongoing. Since popular distributions such as Ubuntu and Fedora may be late in implementing these patches, it is recommended to update the kernel in a timely manner.
Discussions about limited access
TechCrunch raised the question of whether limited access to such a model is enough when AI labs face growing public concerns about protecting digital infrastructure. The discussion follows other OpenAI research, such as a model that unexpectedly hacked a government website during training. While GPT-6 Cyber shows progress, its capabilities and limitations remain important questions for the security community.
Conclusions
GPT‑6 Cyber is a significant step forward in cybersecurity, providing an AI tool that can not only detect but also test real-world software vulnerabilities. The Critical rating highlights that this model has great potential, but also great responsibility. Its access through the Daybreak Red program, strict IP whitelists, and blocking mechanisms show that OpenAI is striving to balance innovation with security. However, as the discussions in the tech media show, it is necessary to constantly monitor how such technologies are used in practice and what measures are in place to avoid unintended consequences.
Future prospects
OpenAI plans to expand GPT-6 Cyber’s reach, but will maintain strict security protocols. This could encourage broader collaboration with security researchers and help identify and remediate critical vulnerabilities more quickly. At the same time, as The Register points out, it is necessary to constantly assess whether the model’s behavior complies with the intended „rules of engagement“ to avoid unexpected violations.
What does this mean for the Lithuanian IT community?
Although GPT-6 Cyber access is currently limited, Lithuanian companies and researchers could leverage this technology through formal OpenAI partnerships or by participating in the Daybreak Red program. This could provide an opportunity to detect local software vulnerabilities early and contribute to the global cybersecurity ecosystem.
