GPT-6 Cyber: OpenAI's New Security AI Explained

Flat illustration of GPT-6 Cyber as a security-focused AI model, with a shield and code symbols in coral, navy and cream

OpenAI has introduced a specialised AI model built for security work, not for general chat. It is called GPT-6 Cyber, and it is designed for authorised vulnerability research and defensive testing.

This is not a model being opened to everyone. It is a narrower tool, and its first results are measured in security tests.

What is GPT-6 Cyber?

GPT-6 Cyber was introduced on 5 November 2026. OpenAI describes it as a specialised model for authorised vulnerability research and defensive testing.

In plain terms, it is meant to help security teams find weaknesses before attackers do. Vulnerability research means looking for flaws in software and reporting them so they can be fixed. That work needs permission, and agreed limits.

OpenAI has also rated its cyber capability as Critical.

What do the test results show?

The early results focus on measurable security tasks. They do not tell us everything about how the model behaves in real work, but they give a clear starting point.

On ExploitGym, GPT-6 Cyber scored 48.5%, compared with 32.1% for GPT-5.6 Cyber. That is a large gap on the same benchmark.

In a web penetration-testing evaluation, GPT-6 Cyber reached 40.3% vulnerability recall, compared with 32.0% for GPT-6 Astra. Recall is about how many of the vulnerabilities present were actually found. A higher score means fewer known weaknesses were missed.

The model was also linked to two larger sets of findings. It found 81 new Firefox vulnerabilities, and it built working tests for 948 Linux kernel vulnerabilities.

A working test helps show that the issue is real and that a fix can be checked. The numbers do not mean the model finds every flaw, or that it replaces human security researchers.

Why is access gated?

Access to GPT-6 Cyber is gated. It is available first through the API and Codex to approved Daybreak Red teams only.

That is a limited first release, not a general launch. If you are not part of one of those approved teams, you should not expect to use GPT-6 Cyber through the API or Codex at this stage.

Security testing needs trust, permission, and oversight. For more on everyday AI use, see How AI Helps You Work Faster Every Day.

Key takeaways

  • GPT-6 Cyber was introduced on 5 November 2026 for authorised vulnerability research and defensive testing.
  • ExploitGym: 48.5%, compared with 32.1% for GPT-5.6 Cyber.
  • Web penetration-testing recall: 40.3%, compared with 32.0% for GPT-6 Astra.
  • Found 81 new Firefox vulnerabilities; built working tests for 948 Linux kernel vulnerabilities.
  • Access is gated to approved Daybreak Red teams only, first through the API and Codex.
  • OpenAI rated its cyber capability Critical.

What does this mean in practice?

For approved defensive teams, GPT-6 Cyber could help with early security work: spotting possible weaknesses and testing them in a controlled way. That could save time on routine checks and help teams focus on harder decisions, such as how serious a flaw is and how it should be fixed.

For everyone else, the direct effect is limited for now. Keep basic habits: keep software updated, follow your security team or provider, and treat security testing as work that needs permission. A model can be more capable in a test and still be limited in who can use it — see The AI Spending Boom Faces Its Reality Check.

GPT-6 Cyber is not a tool for small shops right now. For simple, low-cost starting points, see How Small Shops Can Use AI Without Spending Money.

Conclusion

GPT-6 Cyber is a focused step in AI security work. Its test scores are higher than the models it was compared with, its Firefox and Linux kernel findings are concrete, and its first release is deliberately limited to approved Daybreak Red teams.

Stronger results, a clear defensive purpose, and gated access — that is the story. It is not a general AI launch, and it is not a tool for everyone.

FAQ

What is GPT-6 Cyber?

GPT-6 Cyber is a specialised model introduced by OpenAI on 5 November 2026 for authorised vulnerability research and defensive testing.

Who can use GPT-6 Cyber?

Access is gated. It is available first through the API and Codex to approved Daybreak Red teams only. It is not available to everyone at this stage.

What were its main test results?

It scored 48.5% on ExploitGym, compared with 32.1% for GPT-5.6 Cyber, and reached 40.3% vulnerability recall in a web penetration-testing evaluation, compared with 32.0% for GPT-6 Astra. It also found 81 new Firefox vulnerabilities and built working tests for 948 Linux kernel vulnerabilities.

Sources

Researched with AI assistance and reviewed for accuracy.

Follow on WhatsApp

For more simple AI tips, follow our WhatsApp channel: https://www.whatsapp.com/channel/0029Vb9XzdHDJ6H4TFNAl61r

Comments