---
title: OpenAI releases GPT-6 Cyber for authorized vulnerability research
url: https://www.dataloco.com/en/openai-releases-gpt-6-cyber-for-authorized-vulnerability-research
published: 2026-11-08T18:21:11+00:00
language: en
section: Security
source: https://www.techtalkthai.com/openai-launches-gpt-6-cyber-vulnerability-research-within-scope/
organizations: OpenAI, Daybreak Red, Codex, ExploitGym, RemoteExploitBench, Dragos
publisher: Dataloco
---

# OpenAI releases GPT-6 Cyber for authorized vulnerability research

OpenAI has introduced GPT-6 Cyber, an artificial intelligence model designed specifically for cybersecurity research. The system is trained to identify and confirm vulnerabilities that are genuinely exploitable while strictly operating within boundaries set by researchers. Access to the model is currently available through the API and Codex for approved Daybreak Red customers. The new model achieved a score of 48.5 percent on the ExploitGym benchmark, which tests attacks against specified vulnerabilities on remote targets. This represents an increase from the 32.1 percent score recorded by the previous GPT-5.6 Cyber model. In web penetration testing scenarios where the source code is not visible, GPT-6 Cyber achieved a recall rate of 40.3 percent. This compares to a 32.0 percent recall rate for GPT-6 Astra. On the RemoteExploitBench benchmark, which utilizes vulnerabilities in GitLab, PostgreSQL, and Jenkins, the new model completed tasks successfully 95 percent of the time. In contrast, GPT-6 Astra on Daybreak Blue refused every task in that specific test environment.

Based on these performance results, OpenAI has classified GPT-6 Cyber as Critical in its cybersecurity Preparedness Framework. This is a higher classification than the High level assigned to GPT-5.6 Cyber. To ensure the model remains aligned with safety goals, OpenAI trained the system to stay within authorized scopes, adhere to defined constraints, and report results accurately. In a specific test called ExploitGym Honeypot, which places easily attackable services outside the work scope as traps, GPT-6 Cyber avoided attacking those services 99.93 percent of the time without a deployment safety prompt. This is an improvement over the 94.49 percent avoidance rate of GPT-5.6 Cyber. When the deployment safety prompt is used, the avoidance rate reaches 100 percent.

In practical application, OpenAI research teams used GPT-6 Cyber to discover 81 previously unknown high severity vulnerabilities in Firefox. Mozilla has already issued patches for several of these findings. The model also reproduced 1,724 vulnerabilities in the Linux kernel and created 948 working exploits. Additionally, it reported 304 Critical vulnerabilities to maintainers, including proposed patches for fixes. Security teams at Cisco and CodeRabbit have also begun using the model for vulnerability research and code review. To prevent misuse, OpenAI added monitoring to verify that model actions match user permissions. It enforces IP allowlisting for Daybreak Red API access by default. Customers using Zero Data Retention for general API use are required to use Private Safety Processing. OpenAI also established a Daybreak Red for Critical Infrastructure group, working with Dragos to evaluate and improve safeguards for approved teams testing critical systems. The company plans to expand access to GPT-6 Cyber to all Daybreak Red customers soon and is opening GPT-5.6 Cyber to all Daybreak Red customers.

## This story in other languages

- [日本語](https://www.dataloco.com/ja/openaiaigpt-6-cyber)
- [Русский](https://www.dataloco.com/ru/openai-predstavila-model-gpt-6-cyber-dlia-poiska-uiazvimostei-v-ramkax-zadannyx-ogranicenii)
- [हिन्दी](https://www.dataloco.com/hi/openai-na-bhathayata-anasathhana-ka-le-gpt-6-cyber-madal-lnaca-kaya)
- [Deutsch](https://www.dataloco.com/de/openai-stellt-gpt-6-cyber-fur-die-sichere-schwachstellenforschung-vor)
- [Italiano](https://www.dataloco.com/it/openai-lancia-gpt-6-cyber-per-la-ricerca-di-vulnerabilita)
- [Norsk](https://www.dataloco.com/no/openai-lanserer-gpt-6-cyber-for-forskning-pa-sarbarheter)
- [Polski](https://www.dataloco.com/pl/openai-wprowadza-gpt-6-cyber-do-badan-nad-lukami-w-zabezpieczeniach)
- [Türkçe](https://www.dataloco.com/tr/openai-gpt-6-cyber-modelini-siber-guvenlik-arastirmalari-icin-sinirli-erisimle-yayinladi)
- [עברית](https://www.dataloco.com/he/openai-gpt-6-cyber)
- [Português (Brasil)](https://www.dataloco.com/pt-br/openai-lanca-gpt-6-cyber-para-pesquisa-de-vulnerabilidades-com-foco-em-seguranca)
- [Français](https://www.dataloco.com/fr/openai-lance-gpt-6-cyber-pour-la-recherche-de-vulnerabilites-dans-des-limites-strictes)