OpenAI introduces GPT-6 Cyber for vulnerability research and defense

Higher scores on cybersecurity evaluations

GPT-6 Cyber scored 48.5% on ExploitGym, an evaluation that tests whether models can exploit a specified vulnerability in a remote target. That compares with 32.1% for GPT-5.6 Cyber, which OpenAI describes as a 51% relative improvement.

OpenAI also tested the model on offensive web penetration testing, where models probe web applications without access to their source code. GPT-6 Cyber achieved 40.3% recall of vulnerabilities, compared with 32.0% for GPT-6 Astra.

In a separate evaluation using RemoteExploitBench, GPT-6 Astra with Daybreak Blue refused 100% of tasks, while GPT-6 Cyber completed 95%. The benchmark uses known vulnerabilities in applications including GitLab, PostgreSQL and Jenkins in a controlled environment. OpenAI classifies GPT-6 Cyber as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber.

Training focused on staying within scope

OpenAI says a central focus of GPT-6 Cyber was alignment. The company trained the model to remain within an authorized scope, follow restrictions on its actions and report its work accurately.

In the ExploitGym Honeypot evaluation, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of cases without deployment safety prompts. GPT-5.6 Cyber avoided them in 94.49% of cases. With the prompts, GPT-6 Cyber stayed within bounds in 100% of tests.

The evaluation places models in isolated environments with supporting services that are deliberately easy to exploit. The model must recognize that attacking those services is outside the assigned task. OpenAI says it has also added cyber-specific misalignment monitoring and targeted monitoring for emerging threats.

Findings in Linux and Firefox

OpenAI researchers used GPT-6 Cyber to produce evidence and reproducible tests for 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity. The model also showed that 30 known issues were more severe than previously reported. OpenAI disclosed the findings to Mozilla with proposed patches, and says Mozilla began landing fixes shortly afterward.

For Linux, GPT-6 Cyber reproduced 1,724 vulnerabilities and developed working exploits for 948. OpenAI says this was more than five times the 166 vulnerabilities exploited by GPT-5.6 Sol using the same research harness. The company disclosed 304 critical findings to maintainers with candidate patches and is working to get other vulnerabilities fixed.

OpenAI also says GPT-6 Cyber helped develop and validate exploits for a critical vulnerability in Brotli, a compression library used by Node.js and other software. Security teams at Cisco, Calif, and CodeRabbit are using the model for vulnerability research and code review.

OpenAI plans broader access for defenders

OpenAI says the model is being used in its own security work, including controlled red-teaming campaigns that identified ways to chain vulnerabilities, escape microVMs and execute code on underlying hosts. The findings were used to patch software, tighten network restrictions and report issues to maintainers and vendors.

The company says it wants to expand Daybreak to organizations of every size, independent researchers and open-source maintainers. Mainline models such as GPT-6 Astra support code review, vulnerability triage and patch development. Daybreak Blue is intended to help verified defenders validate vulnerabilities and fixes, while Daybreak Red supports deeper exploit development and authorized security testing with GPT-6 Cyber.