OpenAI GPT-6 Cyber Gets ‘Critical’ Rating

OpenAI launched GPT-6 Cyber on 5 November 2026, a security model built to find software flaws, develop working exploits for them, and help researchers fix them — and it is the first of OpenAI’s cybersecurity models to be rated “Critical” under the company’s Preparedness Framework. The catch is the access list: for now, only a small group of approved customers in OpenAI’s Daybreak Red programme can use it, through the API and Codex. Everyone else gets the scores, not the model.

Quick answer: GPT-6 Cyber is OpenAI's first 'Critical'-rated security model, launched 5 Nov 2026. It scores 48.5% on ExploitGym — but only approved Daybreak Red teams can use it.

That split — a model powerful enough to earn OpenAI’s highest cyber-risk rating, handed only to vetted security teams — is the real story of this launch. The benchmarks are strong, the real-world finds are striking, and the gate around the model is the point.

What Is GPT-6 Cyber?

GPT-6 Cyber is a specialised version of OpenAI’s GPT-6 Astra model, trained specifically for authorised vulnerability research. It is OpenAI’s fourth cybersecurity model this year and the successor to GPT-5.6 Cyber, and it is designed for three jobs: discovering vulnerabilities, developing working exploits that prove a flaw is real, and assisting with security testing and fixes.

It is not a general-purpose chatbot with a security mode. OpenAI built it for defensive researchers — people whose job is to break into systems they have permission to test, so the flaws can be fixed before attackers find them. The model runs through the API and Codex only; no ChatGPT version has been announced.

For context, the GPT-6 family has been arriving in waves: GPT-6.1 Terra, the cheaper computer-use model, launched on 26 October 2026. Cyber is the security specialist of the family, and its launch puts OpenAI in direct competition with Google’s cybersecurity-first Gemini 4 Carbon, which Google unveiled on 28 October 2026.

Why the “Critical” Rating Matters

“Critical” is the top tier of OpenAI’s Preparedness Framework for cybersecurity — and GPT-6 Cyber is the first of OpenAI’s cyber models to reach it. Its predecessor, GPT-5.6 Cyber, was rated “High”. The step up is not a marketing label: under the framework, a Critical rating means the model can build working zero-day exploits with little human help, which is exactly why access is gated.

OpenAI’s own framing, reported by the Indian Express and others on 5 November, is that the rating reflects capability, not intent. The model is meant for authorised work, and the company has wrapped it in the strongest access controls it has applied to any of its models: identity verification, organisational security requirements, API IP allowlisting by default, and cyber-specific monitoring for activity outside a user’s authorisation.

This is also why the launch looks so different from a normal model release. There is no public demo, no free tier, no waitlist page. The announcement is aimed at security teams, not at general users — and that is deliberate.

How Good Is It? The Numbers OpenAI Published

On OpenAI’s own benchmarks, GPT-6 Cyber is a clear step up from both its predecessor and its base model — but every number below comes from OpenAI, and none has been independently reproduced yet.

Benchmark (OpenAI-reported) GPT-6 Cyber Comparison model
ExploitGym (can the model exploit a known flaw?) 48.5% 32.1% (GPT-5.6 Cyber); 42.4% (GPT-6 Astra)
ExploitGym, hardened setting 36.9% 29.7% (GPT-6 Astra)
Web penetration testing (vulnerability recall) 40.3% 32.0% (GPT-6 Astra)
SEC-Bench Pro (pass@1) 87.0% 85.4% (GPT-6 Astra)
RemoteExploitBench (tasks completed) 95% GPT-6 Astra with Daybreak Blue refused all tasks
ExploitGym Honeypot (stayed in scope, no safety prompt) 99.93% 94.49% (GPT-5.6 Cyber)

The ExploitGym headline is the one to remember: 48.5% against 32.1% for GPT-5.6 Cyber, a 51% relative improvement in OpenAI’s own words. But the Honeypot row may matter more in practice. In that test, OpenAI plants deliberately easy-to-attack services outside the authorised scope and measures whether the model goes after them anyway. GPT-6 Cyber stayed in scope 99.93% of the time without any deployment safety prompt — an out-of-scope rate of just 0.07%, down from 5.51% for GPT-5.6 Cyber — and 100% of the time with the prompt added. For a model whose job is to attack systems, restraint is a feature, not a footnote.

One caveat from OpenAI’s own system card, noted in independent coverage: the model scored 100% on the public version of ExploitBench, but the system card itself warns that historical vulnerabilities in that benchmark may have leaked into the training data. Treat that perfect score as contaminated, not as proof.

What It Found in Real Software

Before launch, OpenAI’s researchers used GPT-6 Cyber against real software — and the Firefox results are the headline. The model identified 81 previously unknown vulnerabilities in Mozilla Firefox that Mozilla classified as high severity, and Mozilla has begun shipping fixes, including in its Firefox 159 security advisory (MFSA 2026-118). The model also flagged 30 previously reported Firefox issues that turned out to be more serious than first assessed.

OpenAI also says the model developed working exploits for 948 of 1,724 reproduced Linux kernel vulnerabilities. That kernel pipeline is a separate, detailed story — our sister site OpenAIMaster.com covered the finding itself, including the cost and the patch status, at OpenAI’s AI Found 948 Linux Kernel Flaws — so this piece keeps the figure as one line of supporting evidence for the launch, not the headline.

Taken together, the two campaigns are OpenAI’s argument that the model does more than pass benchmarks: it finds flaws in software millions of people actually use, and supplies reproducible tests and suggested repairs alongside the findings.

Who Can Actually Use It — and Who Cannot

Almost nobody, for now — and that is by design. GPT-6 Cyber is initially available only to selected approved Daybreak Red customers, through the API and Codex. OpenAI says access will expand to all approved Daybreak Red customers over the coming weeks, and GPT-5.6 Cyber is opening to every Daybreak Red customer from the same day.

Access route Who gets it What it allows
Daybreak Red (GPT-6 Cyber) Selected approved customers at launch; all approved customers in the coming weeks Advanced authorised work: exploit development, penetration testing, red teaming
Daybreak Red (GPT-5.6 Cyber) All Daybreak Red customers, from 5 Nov 2026 Previous-generation cyber model, same programme
Daybreak Blue Vetted defensive researchers, broader and less restrictive Vulnerability analysis, defensive research, patch validation (GPT-6 Astra)
General public / ChatGPT No access announced Nothing — no ChatGPT version, no free tier, no public waitlist

Even inside Daybreak Red, access is not automatic: organisations already approved for the programme need separate permission to use GPT-6 Cyber, and must meet requirements including single sign-on, multi-factor authentication and API IP allowlisting. VentureBeat reports, from OpenAI’s developer documentation, API pricing of $20 per million input tokens and $100 per million output tokens, with cached input at $2 per million — far above general-purpose model pricing, and consistent with a tool aimed at funded security teams rather than casual users.

If you are a security researcher who wants in, the route is an application to Daybreak Red describing the specific defensive workflows you plan to use the model for. There is no published timeline for wider availability, and no date for any ChatGPT version.

Our Take: The Gate Is the Story

The most important number in this launch is not 48.5% — it is the size of the access list. OpenAI has built a model it considers capable of critical-level cybersecurity work, and its first move was to restrict it to vetted teams with identity checks, network allowlists and activity monitoring. That is the correct response to a model that can build working exploits — and it is also a quiet admission that the safety question has moved from “can the model do it?” to “who is allowed to ask?”

For most readers, GPT-6 Cyber will never be a tool they touch. Its effect arrives second-hand: flaws found in Firefox and the Linux kernel before attackers find them, and a higher bar for what security teams can automate. The honest test of this launch comes next — whether independent researchers reproduce the ExploitGym gains, and whether the Daybreak Blue defensive tier (where GPT-6 Astra does the fixing work) proves as useful as the Red tier is powerful. Until then, the verdict is: a genuinely capable model, honestly labelled, behind a gate that should stay closed to everyone except the people it was built for.

Desk-researched from OpenAI’s announcement and system card via independent reporting, verified 6 November 2026. Not hands-on tested — the model is restricted to approved Daybreak Red customers, and OpenAIMaster.ai is not one of them.

Frequently Asked Questions

What is GPT-6 Cyber?

GPT-6 Cyber is OpenAI’s specialised cybersecurity model, launched on 5 November 2026. It is a version of GPT-6 Astra trained to find software vulnerabilities, develop working exploits that prove a flaw is real, and help security teams fix them. It is designed for authorised security research only.

Can I use GPT-6 Cyber?

Almost certainly not yet. It is available only to selected approved customers in OpenAI’s Daybreak Red programme, via the API and Codex. There is no ChatGPT version, no free tier and no public waitlist. Security teams can apply to Daybreak Red, describing the defensive work they plan to do.

What does the “Critical” rating mean?

“Critical” is the top cybersecurity tier in OpenAI’s Preparedness Framework. It means OpenAI assesses the model as capable of building working zero-day exploits with little human help. GPT-6 Cyber is the first of OpenAI’s cyber models to reach that rating; GPT-5.6 Cyber was rated “High”.

How much does GPT-6 Cyber cost?

According to VentureBeat, citing OpenAI’s developer documentation, the API price is $20 per million input tokens and $100 per million output tokens, with cached input at $2 per million. That pricing only matters to approved Daybreak Red customers — the model is not on general sale.

Is GPT-6 Cyber dangerous?

It is powerful, which is why access is restricted. OpenAI built in safeguards including cyber-specific monitoring, IP allowlisting and checks on high-risk requests, and its Honeypot test showed the model stayed within authorised scope 99.93% of the time without extra safety prompts. The risk it poses depends entirely on who holds it — which is why the access gate, not the benchmark score, is the story of this launch.

Sources and Methodology

Claims in this article rest on OpenAI’s 5 November 2026 announcement and GPT-6 Cyber system card as reported by independent outlets, cross-checked across at least two sources for every figure, and verified on 6 November 2026. All benchmark figures are OpenAI’s own evaluations and are labelled as such; no independent reproduction was available at publication time.

Leave a Reply

Your email address will not be published. Required fields are marked *