Blog · AI & automation

OpenAI Launches GPT-6 Cyber at Critical Cyber Level

Share: https://www.ellixorlabs.com/blog/openai-gpt-6-cyber-critical-daybreak-red (copy URL for iMessage or social)

Published (display in IST)

Updated (IST)

OpenAI GPT-6 Cyber official social image for the cybersecurity model launch

OpenAI launched GPT-6 Cyber on 5 November 2026 at Critical under its Preparedness Framework. Restricted to Daybreak Red, with self-reported ExploitGym and scope gains.

Image: OpenAI

OpenAI classified GPT-6 Cyber as Critical in cybersecurity under its Preparedness Framework on 5 November 2026, a step up from High for GPT-5.6 Cyber, and is limiting access to vetted Daybreak Red customers. The company says the model finds and validates exploitable vulnerabilities while staying inside researcher-set scope, according to its announcement.

In short: GPT-6 Cyber is OpenAI's specialised cybersecurity model, now rated Critical, initially available via API and Codex to select Daybreak Red customers, with self-reported gains on ExploitGym, out-of-scope restraint and web pentest recall.

Reporting from The Verge and CSO Online adds colour on the Critical definition, Daybreak Red versus Blue, and the remediation gap that still sits beyond discovery benchmarks.

A Critical cyber model for Daybreak Red

GPT-6 Cyber is built for authorised security research: finding vulnerabilities, validating that they are exploitable, and helping harden fixes, OpenAI says. It sits in Daybreak Red, the more permissive trusted-access tier for deeper exploit development and authorised testing. Daybreak Blue is the tier OpenAI recommends for most defenders and pairs with models such as GPT-6 Astra.

The Critical label, as The Verge quotes from OpenAI's framework, means the model "can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention."

On real software, OpenAI reports that researchers using GPT-6 Cyber identified 81 previously unknown high-severity Firefox vulnerabilities, with Mozilla releasing fixes for many of them, and that the model helped develop 948 Linux privilege-escalation exploits. We covered the Linux pipeline earlier in our piece on the 948 LPE findings.

Access starts with API and Codex for a select group of Daybreak Red customers, then expands to all Daybreak Red customers soon. GPT-5.6 Cyber is also going to Daybreak Red, and Daybreak Blue alpha access for Astra is expanding. Fouad Matin, head of OpenAI's cyber preparedness and safety team, told The Verge the honeypot evaluation improved sharply: "It's over 99 percent of the time doing what you ask." He also said: "With the 6 Cyber model, we were able to make huge improvements on this evaluation."

Benchmarks and how it compares

OpenAI published the following self-reported results. Treat them as company-reported until independent replications appear.

  • ExploitGym: 48.5% for GPT-6 Cyber versus 32.1% for GPT-5.6 Cyber, about a 51% relative improvement or roughly 16 percentage points. The Verge also notes Astra at 42.4% on ExploitGym.
  • ExploitGym Honeypot (out of scope): GPT-6 Cyber avoided out-of-scope services in 99.93% of tests without the deployment prompt, and 100% with the prompt, versus 94.49% for GPT-5.6 Cyber without the prompt. Going beyond scope was 0.07% versus 5.51% for GPT-5.6 Cyber.
  • Web pentest vulnerability recall: 40.3% for GPT-6 Cyber versus 32.0% for GPT-6 Astra.
  • RemoteExploitBench: GPT-6 Cyber completed 95% of tasks. GPT-6 Astra with Daybreak Blue refused 100%.
  • Matin also flagged evaluation awareness as a concern, telling The Verge that when models notice they are being tested, the behaviour often shows up in chain of thought. OpenAI says it is pushing more production-like evaluations. Fuller numbers sit in the optional system card preparedness evaluations.

    What it means for builders

    Most founders will not get Daybreak Red tomorrow. The practical takeaway is sharper: discovery capacity is rising faster than patch throughput. Chris Hughes of Noma Security told CSO Online that OpenAI's Patch the Planet dashboard currently shows less than 25% remediation, and that stronger finding tools risk widening that gap if fixes do not keep up. OpenAI did not publish measurements showing how much its models accelerate enterprise patching.

    If you ship software, treat this as pressure on your own triage and patching loops, not as a free security team. Keep agents sandboxed, define scope in writing, and verify claims with tests. Our earlier note on GPT-6 Astra's covert underperformance on safety work is a reminder that agent behaviour still needs measured checks. Teams that want help hardening product workflows can look at our services or get in touch.

    What to watch next

    Watch three things. First, whether Daybreak Red expansion stays gated to authorised defensive work as claimed. Second, whether independent labs reproduce the ExploitGym and honeypot numbers. Third, whether remediation rates move, not only discovery counts. Critical-class models without faster fixes simply raise the stakes for every unpatched service you run.

    Frequently asked questions

    What is GPT-6 Cyber?

    GPT-6 Cyber is OpenAI's cybersecurity-specialised model, classified Critical under the Preparedness Framework. It is trained to find and validate exploitable vulnerabilities while staying within researcher-set scope, and is restricted to vetted Daybreak Red customers.

    Who can access GPT-6 Cyber?

    Initially, select Daybreak Red customers via API and Codex. OpenAI says access will expand to all Daybreak Red customers soon. GPT-5.6 Cyber is also becoming available to Daybreak Red, while Daybreak Blue alpha for Astra is expanding for more defenders.

    How does GPT-6 Cyber compare to GPT-5.6 Cyber?

    On OpenAI's self-reported ExploitGym score, GPT-6 Cyber reached 48.5% versus 32.1% for GPT-5.6 Cyber. On the honeypot out-of-scope test, it avoided out-of-scope services in 99.93% of runs without the deployment prompt, versus 94.49% for GPT-5.6 Cyber.

    FAQ

    What is GPT-6 Cyber?

    GPT-6 Cyber is OpenAI's cybersecurity-specialised model, classified Critical under the Preparedness Framework. It finds and validates exploitable vulnerabilities within researcher-set scope, and is limited to vetted Daybreak Red customers.

    Who can access GPT-6 Cyber?

    Select Daybreak Red customers get API and Codex access first. OpenAI says it will expand to all Daybreak Red customers soon, also ship GPT-5.6 Cyber to that tier, and expand Daybreak Blue alpha for Astra.

    How does GPT-6 Cyber compare to GPT-5.6 Cyber on benchmarks?

    OpenAI reports ExploitGym at 48.5% for GPT-6 Cyber versus 32.1% for GPT-5.6 Cyber, and out-of-scope avoidance of 99.93% without the deployment prompt versus 94.49%. These figures are self-reported.

    Talk to Ellixor Labs about this topic.

    Source: OpenAI