Skip to content
AI Security Wire

Published

- 4 min read

By

OpenAI Releases GPT-6 Cyber, Rated Critical for Cybersecurity

OpenAI GPT-6 Cyber Daybreak Red Preparedness Framework exploit development ExploitGym OT/ICS
img of OpenAI Releases GPT-6 Cyber, Rated Critical for Cybersecurity

OpenAI has published the system card for GPT-6 Cyber, a security-specialized variant of GPT-6 Astra that it classifies as Critical in cybersecurity under its own Preparedness Framework. It is our follow-up to earlier coverage of GPT-5.6-Cyber and the first Daybreak Red launch. Every GPT-6 Cyber figure below comes from OpenAI’s system card unless stated otherwise.

What the system card reports

The card gives GPT-6 Cyber the same preparedness designations as GPT-6 Astra: Critical in cybersecurity, High in biological and chemical capabilities, and below High in AI self-improvement. OpenAI says the extra training focused on tasks such as reproducing crashes, developing working exploits and searching codebases for vulnerabilities.

The headline benchmark figures:

  • ExploitGym: 48.45% intended-exploit success, against 42.43% for GPT-6 Astra. Success requires obtaining the target’s flag through code execution and confirming the exploit used the intended vulnerability.
  • ExploitGym Hardened: 36.94%, against 29.69% for Astra, on targets with protections such as address-space randomization, stack canaries and the V8 sandbox.
  • ExploitBench, Internal Port: a mean 42% arbitrary code-execution success rate across V8 and Chrome targets, against 39% for Astra. The card itself says reliable exploitation of recently disclosed vulnerabilities “remains challenging” for the model.
  • Defensive-use completion rates: the model completed 100% of general security assistance and vulnerability patching tasks, 96% of exploit development tasks and 97% of red-teaming tasks.

What outside testers found

The card says OpenAI collaborated with NIST CAISSI and separately with Irregular. It summarizes only Irregular’s results, which in our reading are mixed:

  • On CyScenarioBench, GPT-6 Cyber solved 11 of 11 long-horizon challenges at least once, against 8 of 11 for Astra, with a 61% average success rate.
  • On FrontierCyber, it solved 90 of 288 challenges, against 96 of 288 for Astra. Irregular observed no successful attacks on fully hardened targets, and neither model solved any of the 15 highest-difficulty challenges.
  • It solved 22 of 22 Atomic challenges.

OpenAI also says GPT-6 Cyber does not beat Astra on general coding: 82.95% on its internal ExpertSWE against 83.24%. Irregular’s overall finding, in the card’s words, was performance “broadly comparable” to Astra’s, with higher success on long-horizon cyber tasks.

Scope and misuse controls

On an ExploitGym Honeypot test, where an auxiliary service is off limits, GPT-6 Cyber exploited the honeypot in 0.07% of graded rollouts without deployment safety instructions. The card lists 5.51% for GPT-5.6 Cyber, 49.5% for GPT-5.6 Sol and 0% for GPT-6 Astra. With the instructions, OpenAI reports an observed exploitation rate of 0%. The card does not give the number of rollouts in the text we read.

The safeguards the card describes:

  • API IP allowlisting is required for Daybreak Red by default, with exceptions for technical constraints.
  • Cyber-specific misalignment monitoring is meant to separate authorized testing from actions outside the user’s scope.
  • Private Safety Processing reviews related interactions together to catch misuse that a single request would not show. API customers using Zero Data Retention are generally required to use it.
  • New checks aim to block high-risk help involving industrial controls, essential services and election systems, and help that could enable serious injury or offensive control of weapons systems. OpenAI says it evaluated these with Dragos in a controlled lab.

Separately, Mozilla’s advisory MFSA 2026-118 for Firefox 159, announced on 27 October, lists 167 CVEs, 109 of them rated high. It credits 41 CVEs, all rated high, to OpenAI reporters: 34 to “OpenAI Codex Security (kylebot)” and 7 to Amy Burnett of OpenAI. The advisory names no model, so it does not tie those reports to GPT-6 Cyber. The credit counts are our tally of the advisory page.

What this means

For defenders, the useful reading is that the gains are incremental and uneven. The card shows a lead over Astra on exploit benchmarks, parity or a deficit elsewhere, and no wins against fully hardened targets in Irregular’s testing. Our assessment is that maintainers should still plan triage capacity for more AI-assisted vulnerability reports, since OpenAI-credited reports already make up a notable share of that Firefox advisory (41 of 167 CVEs), which is our reading of the credits.

The caveat is evidence quality. The benchmarks and the Critical rating are OpenAI’s, and the only external results the card summarizes come from one firm, Irregular. Treat the figures as claims until outside testers publish their own.

References

Frequently Asked Questions

What is GPT-6 Cyber?
According to OpenAI's system card, dated 5 November 2026, it is a cyber-specialized variant of GPT-6 Astra built for authorized security work such as vulnerability discovery, exploit development and web pentesting. The card says it is available only to approved Daybreak Red users and organizations.
How does it compare with GPT-6 Astra?
On ExploitGym the card reports a 48.45% intended-exploit success rate against 42.43% for GPT-6 Astra. The card also reports that it does not outperform Astra on general coding evaluations, and that an external assessment by Irregular found broadly comparable performance with higher success on long-horizon cyber tasks.
Has anyone outside OpenAI tested it?
The card says OpenAI worked with NIST CAISSI and Irregular on external assessments. The Irregular results are summarized in the card. We did not find published results from NIST CAISSI as of 2026-11-09 in the sources we checked.