The field of AI safety has seen a major breakthrough. On November 6, 2026, OpenAI officially released GPT-6Cyber, its most powerful cybersecurity-specific model to date. The model achieved a qualitative leap in core capabilities such as proactive software vulnerability discovery, automated validation, and exploit development, delivering a stunning performance record in real-world testing that shocked the industry.
Stunning real-world performance: Firefox uncovers 81 high-severity vulnerabilities, Linux kernel reproduction capability 5x that of the previous generation
In rigorous real-world scenario tests, GPT-6Cyber demonstrated extremely strong offensive and defensive research capabilities:
Browser security push: In security research targeting the open-source browser Firefox, the model successfully pinpointed and discovered 81 previously undisclosed high-severity vulnerabilities, while also providing detailed testing evidence and reproduction methods. OpenAI has now submitted the relevant vulnerability reports and remediation recommendations to Mozilla, which has quickly released multiple corresponding security fixes based on them.
Deep dive into operating system kernels: In security research tests on the Linux kernel, GPT-6Cyber successfully dynamically reproduced 1,724 vulnerabilities and successfully developed working exploits for 948 of them. By comparison, the previous-generation model GPT-5.6Sol was able to successfully exploit only 166 vulnerabilities under the same test environment, meaning the new model's actual output was more than 5x that of the previous generation.
In addition to specific vulnerability research results, the model scored 48.5% on the professional security benchmark ExploitGym, while the previous-generation GPT-5.6Cyber scored 32.1%, an overall performance improvement of about 51%. This test is mainly used to evaluate whether a model can fully execute an exploitation workflow against a specific vulnerability on a remote target.
Strengthened operational scope boundaries and strict security controls
Along with the model's dramatic upgrade in cybersecurity capabilities, OpenAI has adopted extremely strict compliance and risk-control mechanisms to address its potential risks.
In operational scope boundary control tests, the results showed that even without additional safety prompts deployed, GPT-6Cyber made no attempt at all to attack services outside the task scope in as many as 99.93% of tests; after safety prompts were added, this proportion reached 100%, demonstrating extremely high controllability.
Given its powerful penetration and offensive/defensive potential, OpenAI has officially raised the model's cybersecurity capability rating under its Preparedness Framework to Critical (severe/highest level), while the previous-generation product was rated only High. To prevent misuse of the technology, OpenAI has equipped it with multiple dedicated monitoring, credential protection, and sandbox isolation measures, and has imposed stricter access restrictions on high-risk operations involving critical infrastructure such as power, water treatment facilities, and industrial control systems.
Currently, GPT-6Cyber has not yet been fully opened to the public. The first batch is available only through API and Codex to strictly vetted Daybreak Red customers, and access will gradually expand based on progress in safety evaluations.
