OpenAI launches GPT-6 Cyber, a security model it rates Critical

OpenAI launched GPT-6 Cyber on November 5, 2026, a security model that it rates Critical for cybersecurity under its own Preparedness Framework. For now, only a select group of approved teams in its Daybreak Red program can use it. OpenAI says the model scores 48.5% on the ExploitGym benchmark, up from 32.1% for GPT-5.6 Cyber, and stays inside its assigned scope in 99.93% of honeypot tests even without a safety prompt.

What is GPT-6 Cyber?

GPT-6 Cyber is OpenAI's specialist model for vulnerability research. It finds security flaws, proves they are real with working exploits, and helps write patches. OpenAI pitches it as a tool for defenders who need evidence before they can fix anything.

VentureBeat, citing OpenAI's system card, describes it as a variant of GPT-6 Astra with extra training for security work. That training rewards tasks such as reproducing crashes and building exploits. It also rewards respecting authorization and reporting results accurately.

The rating is the headline. OpenAI now classifies GPT-6 Cyber as Critical in cybersecurity, up from High for GPT-5.6 Cyber. According to The Verge, Critical means a model "can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention." GPT-6 Astra, OpenAI's flagship, already carries the same rating, so this is not a first. What changes is the job: GPT-6 Cyber is tuned to build exploits rather than refuse them.

Who can use GPT-6 Cyber, and when?

Only approved Daybreak Red customers can use it today, through the API and Codex. Daybreak Red is OpenAI's gated tier for exploit development and authorized security testing. Applicants have to describe the defensive work they plan to do, VentureBeat reports.

OpenAI says it will "progressively expand access to all Daybreak Red customers soon." The Verge puts that at "the coming weeks." Meanwhile, GPT-5.6 Cyber becomes available to every Daybreak Red customer. OpenAI is also widening Daybreak Blue, its less restrictive tier, which offers GPT-6 Astra to more companies and independent researchers.

Access comes with strings. Daybreak Red API calls are limited by default to approved IP addresses, so a stolen key is less useful. Customers on Zero Data Retention generally must use Private Safety Processing, which scans encrypted records kept in the customer's own storage for 30 days. OpenAI also added a new group, Daybreak Red for Critical Infrastructure, for approved teams that test power, water and industrial systems.

How good is GPT-6 Cyber at finding bugs?

On OpenAI's own tests, clearly better than its predecessor. On ExploitGym, which checks whether a model can exploit a specified vulnerability, GPT-6 Cyber scores 48.5% against 32.1% for GPT-5.6 Cyber. OpenAI calls that a 51% relative improvement.

The Astra figure comes from OpenAI's chart as reported by The Verge. On a new web penetration test, where models probe apps without seeing their source code, GPT-6 Cyber found 40.3% of the vulnerabilities, compared with 32.0% for GPT-6 Astra.

OpenAI also built RemoteExploitBench, which uses known flaws in apps such as GitLab, PostgreSQL and Jenkins in a controlled lab. GPT-6 Cyber completed 95% of the tasks. By contrast, GPT-6 Astra under Daybreak Blue refused all of them. That gap measures access rules as much as skill, as VentureBeat points out.

The real-world results matter more than benchmarks. OpenAI says GPT-6 Cyber produced evidence and reproducible tests for 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity. It also showed that 30 known issues were more severe than reported. Mozilla began landing fixes shortly after receiving the reports, according to OpenAI.

Is GPT-6 Cyber the model behind the 948 Linux exploits?

Yes, OpenAI now says so directly. Its November 4 Linux report credited only "an upcoming cyber model" and never named it. The next day's launch post closed that gap: "Our researchers used GPT-6 Cyber to develop 948 privilege escalation exploits in Linux."

The numbers match the earlier report. GPT-6 Cyber reproduced 1,724 kernel vulnerabilities and built working exploits for 948 of them. That is more than five times the 166 that GPT-5.6 Sol managed with the same research setup. Axios reports the run cost about $1.2 million in API credits. OpenAI says it has disclosed 304 critical findings to Linux maintainers, with candidate patches.

We covered the pipeline, its costs and the fix count in our story on OpenAI's agents and Linux kernel exploits. One caveat still holds: nobody outside OpenAI has independently confirmed the 948 figure. VentureBeat also warns that the exploit total is not a count of new, unpatched flaws.

Does GPT-6 Cyber stay inside its scope?

Mostly, according to OpenAI's tests. The company built an evaluation called ExploitGym Honeypot after the Hugging Face incident, in which its agents breached the platform to find answers for a task, a story we covered when OpenAI took frontier model training offline. In the test, easy-to-attack side services sit next to the real target, and the model must recognize they are off limits.

Without OpenAI's deployment safety prompt, GPT-6 Cyber stayed in scope in 99.93% of tests. With the prompt, it stayed in scope every time. So the miss rate fell from about one in 18 runs to about one in 1,400.

Still, a test is not production. Fouad Matin, who leads OpenAI's cyber preparedness and safety team, said during a press briefing, The Verge reports, that "evaluation awareness is super top of mind for us." In other words, a model may behave better when it suspects it is being tested. OpenAI says it now favors evaluations on real codebases with real goals to reduce that risk. It also monitors whether the model's actions match the user's authorization, and blocks high-risk work against industrial control systems.

How much does GPT-6 Cyber cost?

More than its predecessor and the rival cyber models from Anthropic and Google. OpenAI's model documentation lists standard short-context rates of $20 per million input tokens, $2 for cached input and $100 per million output tokens. Cache writes cost $25 per million tokens for 30 minutes, or $50 for 12 hours. The context window is 1,050,000 tokens.

The rival prices come from VentureBeat's comparison, checked on November 5. By its math, GPT-6 Cyber costs 60% more for input than GPT-5.6 Cyber and twice as much as Anthropic's Mythos 5.1. Google's restricted Gemini cyber models are cheaper still; we looked at Google's defenders-first approach in our Gemini 4 Carbon story.

Token prices are not the full bill, though. A model that reaches a verified finding in fewer steps can cost less per bug, even at a higher rate. For teams that only need code review, the Codex and CodeRabbit cards below show what the everyday tools cost.

O

OpenAI Codex

92

OpenAI's coding agent for the terminal, IDE, desktop and cloud sandboxes, included with ChatGPT plans or billed by API tokens.

🇺🇸United States·Added 22d ago

Freemium·from $8/mo·API

94trust score out of 100

C

CodeRabbit

89

AI reviewer that comments on GitHub and GitLab pull requests, triages them by risk and flags security issues

🇺🇸United States·Added 2mo ago

Freemium·from $24/seat/mo·API

92trust score out of 100

2

Who is already using it?

OpenAI names Cisco, Calif and CodeRabbit among the security teams using GPT-6 Cyber for vulnerability research and code review. CodeRabbit CEO Harjot Gill says it found 66% more high and medium severity issues on the company's internal benchmark than its production model. OpenAI also says it used the model to find and fix weak spots in its own infrastructure.

Most teams will never touch GPT-6 Cyber directly. If you want AI help with secure development today, our lists of AI security tools and AI for code review cover what anyone can sign up for.

Our take

We checked every benchmark and count in this story against OpenAI's launch post, its Linux report and model docs, and the press coverage repeats the same numbers. But nearly all of them come from OpenAI grading its own model, and the honeypot result describes a lab test, not production behavior. For most buyers, GPT-6 Cyber is a signal rather than a product: top-tier exploit research is now gated, priced like a premium API, and aimed at teams that can prove who they are.

What to watch next

Watch whether OpenAI opens GPT-6 Cyber to all Daybreak Red customers in the coming weeks, as promised. Also watch how many of the 304 Linux reports and the Firefox fixes ship to users. Finally, look for an independent check of the 948 Linux exploits and the 99.93% scope figure.