GPT-6 Cyber: Benchmarks, Daybreak Access & vs GPT-6 Astra
GPT-6 Cyber is OpenAI's security model: 48.5% on ExploitGym, 81 new Firefox bugs, Critical cyber rating. Benchmarks, how to get Daybreak access and how it differs from Astra.
Quick answers
- What is GPT-6 Cyber?
- the Cyber model is OpenAI's cybersecurity model, a variant of GPT-6 Astra trained to find and validate exploitable vulnerabilities while staying inside the scope researchers set. OpenAI announced it on November 5, 2026.
- How good is GPT-6 Cyber?
- It scores 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber, and completed 95% of RemoteExploitBench tasks that GPT-6 Astra refused.
- Can I use GPT-6 Cyber?
- Only through OpenAI's Daybreak program. the Cyber model is available to select Daybreak Red customers, with access expanding over time.
- What has GPT-6 Cyber found?
- 948 Linux privilege escalation exploits and 81 previously unknown high-severity Firefox vulnerabilities, many already fixed by Mozilla.
GPT-6 Cyber is OpenAI's most capable security model, and the first it rates Critical for cyber. OpenAI introduced GPT-6 Cyber on November 5, 2026. It is built on GPT-6 Astra with extra training for vulnerability discovery, exploit development and web pentesting, plus training to stay within the rules of engagement.
This page covers GPT-6 Cyber benchmarks, what it has already found, how to get access through Daybreak and how GPT-6 Cyber compares with GPT-6 Astra.
GPT-6 Cyber at a glance
Item | GPT-6 Cyber |
|---|---|
Maker | OpenAI |
Announced | Nov 5, 2026 |
Base | GPT-6 Astra, cyber post-training |
Access | Select Daybreak Red customers |
Risk rating | Critical (Preparedness Framework); GPT-5.6 Cyber was High |
Early users | Mozilla, Palo Alto Networks, Cisco, Calif, CodeRabbit, SentinelOne, Cognition |
GPT-6 Cyber benchmarks
Test | GPT-6 Cyber | Compared with |
|---|---|---|
ExploitGym | 48.5% | 32.1% (GPT-5.6 Cyber) |
Web pentest recall | 40.3% | 32.0% (GPT-6 Astra) |
RemoteExploitBench | 95% completed | Astra refused 100% |
Honeypot scope test (no prompt) | 99.93% in scope | 94.49% (GPT-5.6 Cyber) |
Honeypot scope test (with prompt) | 100% in scope | — |
The scope result matters as much as the attack numbers. In ExploitGym Honeypot, OpenAI plants easy-to-hack side services next to the real target. GPT-6 Cyber left them alone in 99.93% of runs without a safety prompt and 100% with one. Cognition says the Cyber model scored 23% on its internal vulnerability benchmark, against a previous best of 13%.
What GPT-6 Cyber has found
- Linux: OpenAI researchers used the Cyber model to build 948 privilege escalation exploits.
- Firefox: 81 previously unknown high-severity bugs, despite earlier hardening with frontier models; Mozilla shipped fixes for many.
- OpenAI's own systems: red-team agents chained bugs to escape microVMs on training infrastructure, which OpenAI then patched.
- Customers: Cisco, Calif and CodeRabbit use the Cyber model for vulnerability research and code review; Palo Alto Networks brings it to clients through Unit 42.
How to get GPT-6 Cyber access
the Cyber model is not in ChatGPT or the public API. OpenAI runs access through Daybreak: Daybreak Blue lets verified defenders validate bugs and fixes with fewer refusals using mainline models like GPT-6 Astra, and Daybreak Red adds the Cyber model for exploit development and authorized testing. Organizations apply on OpenAI's Daybreak form; OpenAI says it plans to expand the Cyber model access to companies of every size, independent researchers and open-source maintainers.
GPT-6 Cyber vs GPT-6 Astra
GPT-6 Astra already handles code review, triage and patch work. the Cyber model outperforms Astra on most cyber tasks even when Astra does not refuse, and it completes offensive tasks Astra refuses outright. Outside security, the Cyber model keeps Astra's refusals, so it is not a general "no limits" model.
Let a Sumus AI Employee run your security busywork
Yes, your team can get help with the security work around the model. This post was written by a Sumus AI Employee: it read OpenAI's launch post and system card and built the card above. A Sumus AI Employee can track advisories, summarize new CVEs for your stack and keep the follow-up list in your team chat.
What we're watching
- Wider the Cyber model access for open-source maintainers.
- Independent tests of the Cyber model by outside red teams.
- Anthropic's and Google's answer to a Critical-rated cyber model.
FAQ
When was GPT-6 Cyber released?
OpenAI announced the Cyber model on November 5, 2026.
Is GPT-6 Cyber in ChatGPT?
No. the Cyber model is only for select Daybreak Red customers.
What is Daybreak Red?
OpenAI's program for authorized offensive security work, including exploit development, using the Cyber model.
Why is GPT-6 Cyber rated Critical?
Its exploit skills crossed OpenAI's Critical threshold for cybersecurity in the Preparedness Framework, up from High for GPT-5.6 Cyber.
Are GPT-6 Cyber benchmarks independent?
They come from OpenAI and its partners; the system card has full settings.
Can a Sumus AI Employee help a security team?
Yes. A Sumus AI Employee tracks advisories, writes summaries and keeps follow-ups in your team chat.
Update log
- — Nov 7: First published. GPT-6 Cyber announced Nov 5, 2026 for Daybreak Red customers.