OpenAI releases GPT-6 Cyber model with stronger vulnerability research and alignment safeguards

OpenAI has released GPT-6 Cyber, a specialized model designed for authorized security research. The model improves on its predecessor in finding vulnerabilities and developing exploits while adding stronger controls to keep actions within researcher-defined boundaries.

Benchmark gains and real-world results

On the ExploitGym benchmark, which measures the ability to exploit a specified vulnerability in a remote target, GPT-6 Cyber scored 48.5 percent compared with 32.1 percent for GPT-5.6 Cyber. In a separate web penetration testing evaluation without source code access, it achieved 40.3 percent recall versus 32.0 percent for the general GPT-6 Astra model.

Researchers at OpenAI used the model to identify 81 previously unknown high-severity vulnerabilities in Firefox. Mozilla has already released fixes for many of them. The same work also reproduced 1,724 Linux kernel vulnerabilities and produced working exploits for 948 of them, more than five times the number achieved with the prior model in the same harness. OpenAI disclosed 304 critical findings to Linux maintainers along with candidate patches.

Alignment improvements

A key focus was preventing the model from exceeding task scope. In an ExploitGym Honeypot evaluation designed to tempt the model into attacking supporting services outside the assigned task, GPT-6 Cyber stayed within bounds in 99.93 percent of tests without a deployment prompt and 100 percent with the prompt. This is a substantial improvement over the previous version.

Access and safeguards

GPT-6 Cyber is available only to vetted customers through OpenAI's Daybreak Red program for advanced authorized security testing. It carries a "Critical" rating under the company's Preparedness Framework. OpenAI has added targeted monitoring for misuse, IP allowlisting, and additional blocks for high-risk activity involving critical infrastructure such as industrial control systems.

Security teams at Cisco, Cloudflare, and CodeRabbit are already using the model for vulnerability research and code review. OpenAI itself relied on it during recent work hardening its own frontier model training environment.

The release pairs expanded capability for defenders with tighter controls against misuse. Access remains restricted to approved organizations and researchers.

Sources

• openai.com

Cover photo by Google DeepMind on Pexels, used under the Pexels License.