GPT-6 Cyber finds security holes, but only for approved users

What's new

OpenAI launched GPT-6 Cyber on Nov. 5, 2026, an AI model trained to find and validate vulnerabilities that hackers could exploit. The model must also stay within the scope the security researcher has defined. It is the fourth security model from OpenAI this year, following GPT-5.4 Cyber in April, GPT-5.5 Cyber in June and GPT-5.6 Cyber in August.

It is not for everyone. Access requires approval in the Daybreak Red program, where security teams must document their defensive purpose and, among other things, undergo identity verification, multifactor authentication and monitoring. The model is available only through the API and Codex. A new group, Daybreak Red for Critical Infrastructure, allows approved defenders of critical infrastructure to perform authorized testing of critical systems.

What's clever

The model is meant to be good at staying within bounds. OpenAI researcher Eric Wallace says: "We have worked to ensure that the model is both really good at listening to your instructions about what you should and shouldn't touch in a particular codebase."

According to OpenAI, the model used services outside the permitted scope in 0.07 percent of graded runs without safety instructions, compared with 5.51 percent for GPT-5.6 Cyber and 49.5 percent for GPT-5.6 Sol. With safety instructions, the figure was zero. There are also blocks against high-risk activity targeting critical systems such as industrial control systems, and frequent check-ins during risky actions.

Cheaper or better?

Better, but more expensive. On the ExploitGym test, GPT-6 Cyber scored 48.5 percent versus 32.1 percent for its predecessor, GPT-5.6 Cyber. OpenAI says the model found 81 previously unknown vulnerabilities in Firefox, built exploits (attack code that takes advantage of a vulnerability) for 948 vulnerabilities in the Linux kernel and found a critical flaw in the compression tool Brotli. According to the system card, it completed 100 percent of the tasks in vulnerability patching, 96 percent in exploit development and 97 percent in red teaming (authorized attacks on one's own systems). In Irregular's external test, FrontierCyber, it solved 90 of 288 tasks, but not the 15 hardest. Most of the figures come from OpenAI itself.

The API price is $20 per million input tokens (chunks of text) and $100 per million output tokens. Cached input costs $2. According to VentureBeat, that is 60 percent more for input than GPT-5.6 Cyber and well above Anthropic's Claude Mythos 5.1 and Google's Gemini Cyber models. Specific competitor prices are not given in the sources. There is no standard monthly price, since the model is not sold to individuals.

OpenAI gives the model the highest cyber rating, Critical, in its Preparedness Framework, the same as the regular GPT-6 Astra. OpenAI believes Cyber handles most cyber tasks better than Astra, even when Astra does not refuse the task.

What it's good at

The model is aimed at professional defenders: security teams that need to find holes in their own code, test systems or fix vulnerabilities before others exploit them. It can also be used for red teaming and for testing critical infrastructure by approved users.

For ordinary users and most companies, it is not available. The sources do not say whether the program is open to Danish or European customers, so those interested must apply through Daybreak Red.