OpenAI’s most powerful security model yet is out — but only approved defenders can touch it.

Key Points

  • OpenAI released GPT-6 Cyber on 5 November 2026, a version of its GPT-6 Astra model trained for authorised vulnerability research and defensive security testing.
  • OpenAI rates the model “Critical” for cybersecurity, the top tier of its Preparedness Framework — up from “High” for the earlier GPT-5.6 Cyber.
  • In OpenAI’s own tests, GPT-6 Cyber scored 48.5% on ExploitGym, a benchmark that asks a model to exploit a given flaw in a remote target, against 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra.
  • Researchers used the model to find 81 previously unknown Firefox flaws that Mozilla rated high severity, and to build working exploits for 948 Linux kernel flaws, 588 of them previously unknown.
  • Access starts with a select group of approved Daybreak Red customers through the API and Codex, and will widen to every customer in that tier over the coming weeks.
  • Pricing is set at $20 per million input tokens and $100 per million output tokens, according to OpenAI’s developer documentation.

OpenAI released a new artificial-intelligence model on 5 November that can hunt down software flaws, build working exploits for them and propose fixes — then locked it behind its strictest access controls, so only vetted security teams can use it.

What happened

GPT-6 Cyber is a specialised version of GPT-6 Astra, OpenAI’s general model, with extra training for security investigations. It is built to do the work of a penetration tester: search software for weaknesses, prove a flaw can actually be exploited, and suggest a repair. OpenAI says it also trained the model to stay inside the limits a researcher sets — to test only what it has been authorised to test.

The company put the model through its own research before release. In one exercise, it screened 2,308 candidate findings in the Linux kernel — the core of the operating system that runs most of the world’s servers — reproduced 1,724 vulnerabilities and produced working exploits for 948 distinct local privilege-escalation flaws. Of those, 588 were previously unknown, and OpenAI said it disclosed 304 critical findings to Linux maintainers along with proposed patches. In a separate exercise, the model found 81 previously unknown flaws in the Firefox browser that Mozilla rated high severity; Mozilla has begun releasing fixes.

Independent testers got an early look. “GPT-6 Cyber is the strongest model we’ve tested for security review,” said Harjot Gill, chief executive of CodeRabbit, a code-review company, in OpenAI’s launch announcement.

Access is deliberately narrow. The model is available at first only to a select group of approved customers in Daybreak Red, OpenAI’s access tier for advanced, authorised security work such as penetration testing and exploit development, through the API and Codex. Applicants face identity and eligibility checks, and organisations must use single sign-on and multi-factor authentication, with API requests by default allowed only from approved network addresses. OpenAI said access will expand to all Daybreak Red customers over the coming weeks, and it is also opening the older GPT-5.6 Cyber to every customer in the tier.

Why it matters

The release shows how fast AI is changing the oldest contest in computing: attackers find flaws, defenders patch them. A model that can turn a suspected bug into a working exploit in minutes gives defenders a way to find and fix weaknesses before criminals do — but the same capability in the wrong hands would be a powerful weapon. That is why OpenAI has given GPT-6 Cyber its highest possible cybersecurity rating, “Critical”, and why it is not available to the public at all.

The pricing underlines the point. At $20 per million input tokens and $100 per million output tokens, GPT-6 Cyber costs several times more than general-purpose models, and OpenAI is selling it as a specialist tool for security teams, not a chatbot upgrade. The launch caps a busy week for AI: it follows OpenAI’s acquisition of Pierre, the coding-agent startup, and Tencent’s Hy4 model launch earlier in the week.

There are limits to what the numbers prove. The Linux and Firefox results come from OpenAI’s own research pipeline, not from an independent audit, and a working exploit in a test environment is not the same as a flaw that affects every deployed system. OpenAI’s published account does not mean all 948 patches have been merged into the Linux kernel. What the results do show is that the model can do, at scale, work that used to take skilled researchers days.

What happens next

Three things to watch. First, the rollout: whether OpenAI keeps to its promise to widen access to all Daybreak Red customers over the coming weeks, and how many of the more than 2,000 organisations already enrolled in the programme take it up. Second, the patch pipeline: how many of the hundreds of Linux findings OpenAI reported are confirmed and fixed upstream, and how quickly Mozilla ships the rest of its Firefox fixes. Third, the wider race: rival labs are building their own security models, and the question now is whether the industry can keep the most capable ones in defenders’ hands without slowing the defensive work they were built for.

Sources: OpenAI, “GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents”, 5 November 2026; Reuters, “OpenAI launches GPT-6 Cyber with ‘Critical’ cybersecurity rating”, 5 November 2026; VentureBeat, 5 November 2026. The CodeRabbit quote is from OpenAI’s launch announcement.

Published: 6 November 2026, Kuala Lumpur time.