OpenAI’s GPT-6 Cyber finds more vulnerabilities, but its scope controls face the real test

OpenAI has launched GPT-6 Cyber, a cybersecurity model it says is better at finding vulnerabilities and staying within authorised testing boundaries. The model is available to selected Daybreak Red customers, while its “Critical” rating signals that OpenAI considers its capabilities unusually powerful.

OpenAI Watch analysis

What happened

The model is designed for cybersecurity work, including vulnerability research and testing. The Verge reports that OpenAI’s ExploitGym evaluation gave GPT-6 Cyber a score of 48.5 per cent, compared with 32.1 per cent for GPT-5.6 Cyber. OpenAI also says the model helped identify 81 previously unknown vulnerabilities in Firefox.

OpenAI says a separate evaluation found the model avoided exploiting out-of-scope services more than 99 per cent of the time. The Register reports that 0.07 per cent of cases in that exercise went beyond the intended scope. Those figures describe tests, not a guarantee about behaviour in every real-world engagement. OpenAI has classified the model as “Critical” for cybersecurity under its Preparedness Framework. It is initially available to selected Daybreak Red customers doing authorised defensive security work, with wider access to that customer group planned in the coming weeks.

Why it matters

The same capabilities that help security teams find and test weaknesses can also be misused. GPT-6 Cyber’s launch puts two questions side by side: how much more effective is it at vulnerability work, and how reliably can it stick to the boundaries set by its operator?

That matters beyond one model. Security teams need useful tools for finding flaws, but organisations also need evidence that a system will respect the scope of a test. A strong result in a controlled evaluation is relevant, but it is not the same as proving safe performance across varied live environments.

Our read

This is a substantial cybersecurity release, not just a fresh model name on a familiar box. The reported benchmark result and Firefox work make the offensive capability concrete; the “Critical” classification makes the stakes plain. OpenAI’s scope-adherence figures are worth noting, but they should be read as test results, not a safety certificate. For now, access is limited to approved customers, which gives the company a chance to show how the model performs in authorised work before broadening availability.

What to watch

• Whether OpenAI expands access to all Daybreak Red customers as planned.
• What independent researchers find about the model’s ability to stay within authorised scope.
• Whether OpenAI publishes more detail about the tests and the cases that exceeded their intended boundaries.

Discussion spark: Should a cybersecurity model rated “Critical” be made available to more defenders on the strength of controlled scope tests, or should wider access wait for evidence from real-world use?

OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.