#

Key takeaways

• OpenAI introduced GPT-6 Cyber for approved security teams.

• Access is limited at launch to a select group of approved Daybreak Red customers.

• OpenAI classifies the model as Critical in cybersecurity under its Preparedness Framework.

• The company reports stronger evaluation results than GPT-5.6 Cyber.

• OpenAI says researchers used the model in approved security research and code review.

What OpenAI announced

OpenAI introduced GPT-6 Cyber as a model designed to help approved security teams find, validate, and fix software vulnerabilities within an authorized scope. The company says the launch is restricted. At first, access through the API and Codex is limited to a select group of approved Daybreak Red customers.

OpenAI also says it plans to expand access to other Daybreak Red customers. In the same announcement, it says the previous GPT-5.6 Cyber model is now available to all customers in that tier. That creates a staged rollout, with the newer model reserved for a narrower group at launch.

How OpenAI positions the model

OpenAI classifies GPT-6 Cyber as Critical in cybersecurity under its Preparedness Framework. It says this is a higher classification than the High rating it gave GPT-5.6 Cyber. That comparison is part of the company's own framework and should be read in that context.

The announcement frames the model as a tool for authorized work. It is meant to support security teams that already have permission to operate within a defined scope. The source material ties the release to approved cybersecurity research and code review.

Evaluation results reported by OpenAI

OpenAI says GPT-6 Cyber scored 48.5 percent on ExploitGym. In that evaluation, the model must exploit a specified vulnerability. The company says GPT-5.6 Cyber scored 32.1 percent on the same benchmark.

OpenAI also reports results from a separate authorized-scope evaluation. In those tests, the model avoided exploiting out-of-scope services in 99.93 percent of cases without its deployment prompt. With the prompt, it avoided out-of-scope services in 100 percent of tests. OpenAI says these figures describe its own tests, and they are separate from the model's vulnerability discovery results.

Research use and supporting evidence

OpenAI says researchers used the model to develop 948 Linux privilege escalation exploits. It also says the model helped identify 81 previously unknown high-severity Firefox vulnerabilities. The announcement says the model also helped develop candidate patches.

OpenAI adds that Mozilla has released fixes for many of the Firefox issues. The source material presents these research results as supporting evidence for the restricted model release. It also names approved cybersecurity research and code review by teams at Cisco, Calif, and CodeRabbit.

Operational considerations

The announcement suggests a controlled deployment model. Access is limited, and the company says it will expand access later. That structure matters because the model is intended for authorized security work, not open use.

The separate evaluation results also matter. OpenAI distinguishes between exploit performance and authorized-scope behavior. That distinction helps readers understand what each metric measures. It also shows that the company is presenting more than one kind of evidence for the release.

Why this release matters

GPT-6 Cyber is presented as a cybersecurity-focused model with tighter access controls and stronger reported results than its predecessor. OpenAI's announcement emphasizes both capability and scope control. It also links the launch to real research activity, including Linux exploit development and Firefox vulnerability discovery.

For readers, the main point is simple. OpenAI is positioning GPT-6 Cyber as a restricted tool for approved teams, with performance claims backed by company-run evaluations and research examples.