OpenAI released GPT-6 Cyber on Thursday, November 5, 2026, and gave it the highest cybersecurity rating in its own safety system. The model is a cyber-specialized version of GPT-6 Astra, built for authorized vulnerability discovery, exploit development, and web penetration testing. OpenAI classifies it as Critical for cybersecurity under its Preparedness Framework, which is the top tier in that framework. The predecessor, GPT-5.6 Cyber, was rated High, so this is a clear step up in how OpenAI judges the power of its own security model. Access is not open to the public. At launch, only a select group of approved Daybreak Red customers can use it through the API and Codex, and OpenAI says broader access is coming soon without giving a firm date or price.
Quick Facts
| Fact | Detail |
|---|---|
| Model name | GPT-6 Cyber |
| Release date | Thursday, November 5, 2026 |
| Base model | GPT-6 Astra, cyber-specialized variant |
| Cybersecurity rating | Critical, the top tier in OpenAI Preparedness Framework |
| Previous model rating | GPT-5.6 Cyber was rated High |
| Linux kernel exploits developed | 948 distinct local privilege escalation flaws, as reported by OpenAI |
| Linux kernel vulnerabilities reproduced | 1,724, as reported by OpenAI |
| Novel Firefox vulnerabilities found | 81, classified high severity by Mozilla, as reported by OpenAI |
| Critical findings sent to maintainers | 304, with proposed patches, as reported by OpenAI |
| ExploitGym score | 48.5 percent, versus 32.1 percent for GPT-5.6 Cyber, as reported by OpenAI |
| Access at launch | Select approved Daybreak Red customers only, through the API and Codex |
| Access controls | Identity verification, single sign-on, multi-factor authentication, API IP allowlisting |
What GPT-6 Cyber Is and Why It Exists
GPT-6 Cyber is not a general chatbot with a few extra security tricks. OpenAI describes it as a model trained for authorized offensive security work, which means finding weak points in software, building working proof-of-concept exploits in a controlled setting, and testing websites for flaws before attackers can use them. It comes from GPT-6 Astra, the broader GPT-6 model, but its training and tuning focus on cyber tasks.
The idea behind a model like this is simple to state and hard to do well. Security teams need to find bugs faster than attackers do. A tool that can scan code, trace how a flaw could be used, and show a working exploit helps defenders fix the real problems first.
OpenAI has been moving in this direction all year. GPT-6 Cyber is its fourth cyber-specific model in 2026, after GPT-5.4 Cyber in April, GPT-5.5 Cyber in June, and GPT-5.6 Cyber in August. Each release has pushed capability further, and this one is the first to reach the Critical tier in OpenAI own framework.
For readers who follow OpenAI wider plans, the timing is also worth noting. You can read about the recent funding round in our report on OpenAI 11 billion funding, and about the infrastructure debate in our article on OpenAI Sarah Friar data centers. GPT-6 Cyber sits inside that larger story, as a specialized product aimed at a narrow, high-skill audience.
What the Critical Rating Really Means
The word Critical can sound alarming, and it is meant to be taken seriously, but it helps to be clear about what it is and what it is not. The Critical rating comes from OpenAI own Preparedness Framework. It is the top tier in that framework for cybersecurity. It is not a rating from a government regulator, an independent lab, or an industry standards body.
The jump from High to Critical also tells a story about progress. GPT-5.6 Cyber was rated High. GPT-6 Cyber is rated Critical. OpenAI is saying the new model is meaningfully more capable at finding and exploiting flaws than the model it replaces.
It is also worth keeping the language plain. A Critical rating does not mean the model is loose in the wild, and it does not mean every computer is suddenly at risk from it. It means a powerful security tool exists, and OpenAI has decided to keep it inside a controlled program while it learns how it is used in real work.
The Numbers OpenAI Reported
OpenAI shared a set of results to support the Critical rating, and every number in this section is OpenAI own vendor-reported figure, not an independent benchmark. That caveat should stay in mind while reading the results.
The headline result concerns the Linux kernel, the core of the Linux operating system. OpenAI says GPT-6 Cyber reproduced 1,724 Linux kernel vulnerabilities and developed working exploits for 948 distinct local privilege escalation flaws. A local privilege escalation flaw lets an attacker who already has limited access on a machine gain higher privileges on that same machine. OpenAI compared this with 166 working exploits for GPT-5.6 Sol in an earlier scan using the same research harness.
The browser results are also striking. OpenAI says the model found 81 novel Firefox vulnerabilities that Mozilla classified as high severity. Novel here means the flaws were not already known. The company also says it sent 304 critical findings to maintainers with proposed patches, which points to responsible disclosure rather than simply collecting flaws.
On controlled evaluations, OpenAI reports 48.5 percent for GPT-6 Cyber on ExploitGym, versus 32.1 percent for GPT-5.6 Cyber. ExploitGym is a controlled evaluation of exploiting a specified vulnerability in a remote target. On a separate internal black-box web evaluation, OpenAI reports 40.3 percent vulnerability recall, versus 32.0 percent for GPT-6 Astra. On RemoteExploitBench, OpenAI reports 95 percent completion.
One more result may matter as much as the headline scores. In a scope-following evaluation on a simulated honeypot, GPT-6 Cyber strayed out of scope in 0.07 percent of graded runs, versus 5.51 percent for GPT-5.6 Cyber. Staying in scope is vital in security testing, because a tester that wanders into systems it was not asked to test can cause harm and break trust.
None of these numbers should be read as a final verdict. They are OpenAI own claims, made to explain a rating OpenAI itself assigned. The pattern is consistent, though: across kernel work, browser work, web testing, and scope control, OpenAI reports a clear improvement over the previous generation.
Who Can Use It and How Access Works
At launch, GPT-6 Cyber is restricted to a select group of approved Daybreak Red customers, through the API and Codex. Daybreak Red is OpenAI access tier for authorized offensive security work. That means the model is aimed at professional security teams doing approved testing, not at general users, students, or hobbyists.
Access requires identity verification. The controls OpenAI lists include single sign-on, multi-factor authentication, and API IP allowlisting. In plain terms, OpenAI wants to know who is using the model and to limit API use to approved network addresses.
OpenAI says broader Daybreak Red access is coming soon, but it gave no firm date or price. That leaves a lot unanswered for teams that want in. There is no public price list to compare and no date to plan around.
The restricted launch also reflects the safeguards OpenAI says it added. The company says it added protections for industrial controls, essential services, election systems, and other high-risk targets. Those are the kinds of systems where a mistake or misuse could cause serious harm.
The wider AI industry is watching this kind of launch closely, because capability and control are now moving together. Our report on Anthropic IPO valuation risk looks at how the market is judging one of OpenAI main rivals, and it helps to place GPT-6 Cyber in a competitive field where security capability is becoming a selling point as well as a safety question.
What This Means for Security Teams and Everyone Else
For professional security teams, GPT-6 Cyber could change how much testing they can do and how fast they can do it. A model that finds more flaws and builds working exploits in a controlled setting can help a small team cover more code and give clients clearer proof that a flaw is real.
For software maintainers, the news is mixed but mostly useful. Finding 948 kernel exploits sounds alarming, but OpenAI says it sent 304 critical findings to maintainers with proposed patches. That is the responsible path: find the flaw, tell the people who can fix it, and offer a fix.
For everyday users, the direct effect is limited for now. You cannot sign up for GPT-6 Cyber, and you do not need to. The indirect effect matters more. If security teams can find and fix flaws faster, the software you use every day should get safer over time.
There is also a broader question that this release raises for the whole industry. As AI models get better at finding flaws, the gap between defenders and attackers may narrow or widen depending on who gets access first. OpenAI answer, for now, is to give the tool to approved defenders and to keep it out of general use.
Frequently Asked Questions
What is GPT-6 Cyber?
GPT-6 Cyber is a cyber-specialized variant of GPT-6 Astra, released by OpenAI on November 5, 2026. It is trained for authorized vulnerability discovery, exploit development, and web penetration testing. It is a professional security tool, not a general-purpose chatbot for public use.
Why did GPT-6 Cyber get a Critical rating?
OpenAI classifies GPT-6 Cyber as Critical for cybersecurity under its own Preparedness Framework, which is the top tier in that framework. The predecessor, GPT-5.6 Cyber, was rated High. The rating reflects OpenAI own judgment that the new model is much more capable at cyber tasks. It is OpenAI own category, not a regulator rating.
Can anyone use GPT-6 Cyber?
No. At launch, access is restricted to a select group of approved Daybreak Red customers, through the API and Codex. Access requires identity verification, and controls include single sign-on, multi-factor authentication, and API IP allowlisting. OpenAI says broader Daybreak Red access is coming soon but gave no firm date or price.
How many Linux kernel exploits did GPT-6 Cyber find?
OpenAI says the model reproduced 1,724 Linux kernel vulnerabilities and developed working exploits for 948 distinct local privilege escalation flaws, compared with 166 for GPT-5.6 Sol in an earlier scan using the same research harness. These are OpenAI own reported numbers, not independent benchmark results.
Sources
Final Thoughts
GPT-6 Cyber is a clear sign of where AI security tools are heading. OpenAI reports a large jump in capability, from 166 kernel exploits for an earlier model to 948 for GPT-6 Cyber, plus 81 novel Firefox flaws and stronger scores on controlled tests. At the same time, OpenAI has put the model behind its strictest access controls, with identity checks, multi-factor authentication, and IP allowlisting for a small group of approved security teams.
The honest way to read this release is to hold both facts together. The capability is real enough that OpenAI gave it a Critical rating, and the controls are real enough that most people will never touch the model directly. The numbers still need independent checking, because every benchmark here is OpenAI own report. If the results hold up in wider use, defenders will have a stronger tool for finding flaws before attackers do.
Featured image: AI-generated editorial illustration for this story.
