On 5 November OpenAI released GPT-6 Cyber, a cybersecurity-tuned model it classifies as 'Critical', the top cyber tier of its Preparedness Framework. It says the model produced working proof-of-concept code for 81 Firefox and 948 Linux-kernel vulnerabilities, and access is initially limited to approved 'Daybreak Red' customers cleared for offensive work. The release pairs stronger offensive capability with controls meant to keep the model within authorised scope.
The capability figures and the 'Critical' rating
OpenAI calls it a clear step up from GPT-5.6 Cyber that outperforms the general-purpose GPT-6 Astra on most cyber tasks. It says the model generated 81 previously unknown high-severity Firefox flaws and working exploits for 948 Linux-kernel vulnerabilities, including 588 previously unknown, all reported to maintainers with fixes under way. 'Critical' in the framework denotes a model that can find and develop zero-day exploits in real-world critical systems and run end-to-end attacks.
On the ExploitGym evaluation it scored 48.5%, up from the prior model's 32.1%. On a separate remote-attack benchmark the general-purpose Astra refused every task while GPT-6 Cyber completed 95%. These are the company's own stated evaluations, with no independent verification shown, so the numbers indicate direction rather than confirmed capability.
Gated access and the controls against overreach
Access is initially confined to a subset of Daybreak Red customers cleared for offensive work, used through the API and Codex. OpenAI recommends Daybreak Blue for defensive tasks and set up a group for critical-infrastructure defenders. It says it added identity and account-security requirements, more frequent check-ins to OpenAI on high-risk actions, and monitoring for misalignment, and trained the model to stay within authorised scope, respect restrictions and report results accurately.
In tests without the deployment safety prompt the model left out-of-scope services alone 99.93% of the time, up from the prior model's 94.49%; with the prompt it stayed in bounds every time. OpenAI has not disclosed what happened in the remaining 0.07%, and says it strengthened safeguards against attacks on industrial control systems.
Sources
Primary sources
- 1
Introducing GPT-6 Cyber
OpenAI · External site
Reporting
- 2
OpenAI releases GPT-6 Cyber, and promises this one stays within 'rules of engagement'
Source checked 6 Nov 2026
The Register · External site - 3
OpenAI debuts GPT-6 Cyber, touting better performance — and safety guardrails
Source checked 6 Nov 2026
CyberScoop · External site - 4
OpenAI debuts new cybersecurity-focused AI model amid growing alarm about AI-driven cyber attacks
Source checked 6 Nov 2026
Fortune · External site
From developments to evidence and analysis
Background
No explicit relationship is recorded.
Evidence
No explicit relationship is recorded.
Follow-ups / later versions
No explicit relationship is recorded.
Related questions
Cite this article
AI & Security Observatory（2026-11-05）「OpenAI、サイバー特化「GPT-6 Cyber」を公開 能力枠組み最高位「クリティカル」、攻撃用途は承認制に」『AI & Security Observatory』、第1.0版、最終更新2026-11-06。https://ai-security-observatory.org/ja/brief/brief-20261106-01/
AI & Security Observatory (2026-11-05). OpenAI releases cyber-tuned GPT-6 Cyber, rated 'Critical', limited to vetted offensive users . AI & Security Observatory, version 1.0. Updated 2026-11-06. https://ai-security-observatory.org/en/brief/brief-20261106-01/
Version 1.0 · Publication date 5 Nov 2026 · Updated 6 Nov 2026
Revision history
- Earlier revisions have not been recorded.