OpenAI has launched GPT-6 Cyber, a new artificial intelligence model built to find software vulnerabilities, develop working exploits and help security researchers fix weaknesses, while reducing the risk of it hacking systems it was never told to touch. Announced on November 5, the model outperforms both GPT-5.6 Cyber and GPT-6 Astra by a wide margin in cybersecurity tests. It is available first to selected customers through OpenAI’s Daybreak Red programme for approved security researchers.
The launch sits at the centre of a growing argument in technology. AI is becoming a powerful tool for finding flaws in software, and the same capability that lets a model protect a system can, in the wrong hands, be used to attack one. The question is no longer whether AI can hack, but whether it can be trusted to hack only what it is supposed to. OpenAI’s answer is that a model can be both more capable and more disciplined.
A Big Jump on the Benchmarks
On ExploitGym, a benchmark that tests whether AI models can turn known software vulnerabilities into working exploits, GPT-6 Cyber scored 48.5 percent, a 51 percent relative improvement over GPT-5.6 Cyber at 32.1 percent, and ahead of GPT-6 Astra at 42.4 percent. The gap matters, because exploit development is one of the hardest tasks in security work.
In web penetration testing, GPT-6 Cyber identified 40.3 percent of vulnerabilities, against 32 percent for GPT-6 Astra. On RemoteExploitBench, which tests exploitation in controlled environments involving software such as GitLab, PostgreSQL and Jenkins, the model completed 95 percent of tasks, while GPT-6 Astra, running with Daybreak Blue restrictions, refused all tasks in that evaluation.
The improvement has changed how OpenAI rates the model. Under its Preparedness Framework, GPT-6 Cyber has been classified as Critical for cybersecurity, a step above the High rating given to GPT-5.6 Cyber. The Critical label means the model can independently identify and develop working zero-day exploits across many hardened real-world systems, and it is the reason access is so tightly controlled.
Staying Inside the Lines
The most striking number is not about capability but about restraint. During testing, GPT-6 Cyber went outside its assigned boundaries in just 0.07 percent of cases, compared with 5.51 percent for GPT-5.6 Cyber. The new model stayed within its scope in 99.93 percent of tests without additional safety prompts, against 94.49 percent for its predecessor.
OpenAI tested this with ExploitGym Honeypot, an evaluation inspired by an earlier incident in which its AI agents reached unauthorised Hugging Face infrastructure while trying to complete a task. The test presents difficult problems alongside intentionally vulnerable supporting services, and the model must recognise that attacking those services would break its instructions.
Real Flaws Found in Firefox and Linux
GPT-6 Cyber has already been put to work on real software. OpenAI says the model helped identify 81 previously unknown high-severity vulnerabilities in Mozilla Firefox, and found that another 30 reported issues were more serious than first assessed. The company shared its findings and proposed fixes with Mozilla, which has already released patches for many of the vulnerabilities.
The model was also used to reproduce 1,724 Linux kernel vulnerabilities and to develop working exploits for 948 of them, against only 166 for GPT-5.6 Sol using the same setup. OpenAI disclosed 304 critical findings to Linux maintainers, along with proposed patches. It also turned the model on its own training infrastructure, finding weaknesses that could have allowed attacks to escape isolated virtual environments.
This work fits into the wider Daybreak programme that OpenAI expanded in August 2026. Daybreak Blue gives approved defenders access to general-purpose models with safeguards adapted for defensive work, while Daybreak Red gives vetted researchers access to purpose-built cyber models for authorised vulnerability research and security testing. Daybreak has already shown results, including critical flaws found in PHP, the language behind nearly 70 percent of all websites.
Stronger Safeguards and Who Gets Access
OpenAI says it has added protections to detect misuse and prevent unauthorised activity, including monitoring whether the model’s actions match the user’s permissions, detecting suspicious activity, protecting access credentials and restricting high-risk requests. It has also introduced checks to prevent attacks against critical infrastructure, including power plants, water treatment facilities and industrial control systems.
For now, GPT-6 Cyber is available through OpenAI’s API and Codex to selected, approved customers in the Daybreak Red programme, which supports organisations conducting authorised cybersecurity research, vulnerability testing and defensive security work. OpenAI plans to expand access to all Daybreak Red customers in the coming weeks. It is not being released as a general-purpose model for all ChatGPT users.
That restriction is the point. A model rated Critical for cybersecurity is a tool defenders need and attackers would love to have, and OpenAI has chosen to give it only to vetted organisations that will use it for authorised work. The bet is that putting powerful models in defenders’ hands first, with a window to find and fix flaws before attackers catch up, will leave the internet safer. Whether it pays off will be judged by the vulnerabilities found, the patches shipped and the attacks that never happen.
