---
title: "OpenAI releases GPT-6 Cyber to approved security teams, with access still limited"
description: "OpenAI releases GPT-6 Cyber to selected Daybreak Red customers. Compare five benchmarks, external evaluation caveats and the access requirements for security teams."
url: "https://promptszone.com/openai-gpt-6-cyber-release-access-benchmark-comparison"
published: "2026-11-05T19:00:05Z"
categories: ["New Models"]
---

# OpenAI releases GPT-6 Cyber to approved security teams, with access still limited

GPT-6 Cyber begins a restricted rollout for approved security teams. Its benchmark gains need to be read alongside access requirements and mixed external results.

![Original five-row scorecard comparing GPT-6 Cyber with GPT-6 Astra on OpenAI-reported cybersecurity evaluations, with the official OpenAI provider wordmark.](https://promptszone.com/assets/uploads/a1978a60acd01ed2979a5f440027af36.png)

**Summary:**

OpenAI’s cyber-specialized model is initially available to selected Daybreak Red customers. A five-benchmark comparison shows gains over Astra, while external evaluations add important limits.

OpenAI introduced GPT-6 Cyber on November 5, starting with a select group of approved Daybreak Red customers through the API and Codex. Broader access for Daybreak Red customers is planned, rather than available to everyone at launch.

The release focuses on authorized security research, including finding vulnerabilities and validating fixes. OpenAI reports stronger performance than GPT-5.6 Cyber on ExploitGym and improved adherence to assigned boundaries. These are developer claims, not a guarantee that an autonomous security agent will remain within scope in every environment. [Read the official announcement](https://openai.com/index/introducing-gpt-6-cyber/).

## Access requires more than a Daybreak Red account

OpenAI’s [current Daybreak guidance](https://help.openai.com/en/articles/20001258-openai-daybreak-trusted-access-for-cyber-overview) says GPT-6 Cyber needs separate model approval. Enabling Daybreak Red does not itself unlock the model. The normal Red application route is currently for approved business and enterprise organizations; individual access can exist by exception.

Daybreak is restricted to approved internal work on systems an organization owns or is explicitly authorized to test. Customer-facing traffic and onward access for third parties are excluded. Procurement teams should therefore verify eligibility, approved users, data handling and permitted workflows before treating the announcement as a deployable service.

## Five benchmarks, one carefully bounded comparison

The [November 5 system card](https://deploymentsafety.openai.com/gpt-6-cyber/gpt-6-cyber.pdf) reports the following at the highest tested reasoning budget. All are OpenAI-reported. Higher is better within each row; bold marks the best numeric result. Different metrics must not be averaged into an overall score.

| Evaluation and metric | GPT-6 Astra | GPT-6 Cyber |
| --- | --- | --- |
| ExploitBench internal port, June–August 2026: mean success | 39% | **42%** |
| SEC-Bench Pro: pass@1 | 85.38% | **87.02%** |
| SRE-Bench: pass@1 | 87.98% | **92.84%** |
| ExploitGym: intended-exploit success | 42.43% | **48.45%** |
| ExploitGym Hardened: intended-exploit success | 29.69% | **36.94%** |

This is a same-provider comparison, not a cross-vendor leaderboard. Comparable current scores for leading rival models were not established in the reviewed sources. Historical system-card scores should not be mixed into this table.

## External results add an important qualification

Irregular’s updated evaluations, summarized in the system card, show a mixed picture: Cyber solved 11 of 11 CyScenarioBench challenges at least once versus Astra’s 8, but 90 of 288 FrontierCyber challenges versus Astra’s 96. Neither solved the 15 hardest FrontierCyber challenges. These external evaluations are reported through OpenAI’s card, not independently reproduced here.

## What a useful pilot should establish

For an eligible security team, the useful question is whether the model improves the work that reaches a maintainer or reviewer. A pilot can track validated findings, false positives, reviewer time, patch quality and cost on an explicitly authorized test set. It should also test whether the agent reports limitations clearly and stops when its permitted scope ends.

Benchmark gains do not remove the need for human review or permission to test a system. Keep evaluation environments isolated, document the allowed targets and actions, and review results before changing production software. This article reports the release and its evidence; it does not provide exploitation instructions.

Image: original PromptsZone chart from OpenAI’s November 5 system card, not a product screenshot. The unchanged official OpenAI provider wordmark is used as a fallback because no distinct GPT-6 Cyber logo was identified. OpenAI owns the mark; no endorsement is implied. [Brand guidance](https://openai.com/brand/).