OpenAI launches GPT-6 Cyber model with critical security rating

OpenAI has launched GPT-6 Cyber, a specialized artificial intelligence model built to discover software vulnerabilities, claiming the system is significantly less likely to go off script or breach unauthorized networks while conducting security research.

The model is initially rolling out to select members of OpenAI's Daybreak Red program, which provides tools to vetted cybersecurity professionals and organisations doing authorized defensive research, before expanding to all program participants in the coming weeks.

OpenAI has assigned GPT-6 Cyber a "Critical" designation under its internal Preparedness Framework, up from the "High" rating given to its predecessor, GPT-5.6 Cyber. The classification means the model can independently identify and build working zero-day exploits across hardened real-world systems.

During safety testing on ExploitGym HoneyPot—an evaluation designed to measure whether autonomous agents stray into unauthorized systems—GPT-6 Cyber went out of bounds less than 1 per cent of the time, compared to roughly 5 per cent for GPT-5.6 Cyber.

Software vulnerability testing and safety guardrails

Fouad Matin, head of OpenAI's cyber preparedness and safety team, said the lab prioritises real-world testing environments to evaluate whether models recognize they are being tested. "With the 6 Cyber model, we were able to make huge improvements on this evaluation," Matin said. "It's over 99 percent of the time doing what you ask."

In early testing, OpenAI said the model helped identify 81 previously unknown vulnerabilities in the Firefox web browser and was used to harden the company's internal AI training infrastructure. The company also disclosed that an automated AI pipeline using the model scanned and prepared patches for 948 potential flaws in the Linux kernel, spending $1.2 million in compute resources on the Linux audit.

Independent security researchers have already demonstrated the capability of OpenAI's underlying technology. Max Justicz, chief technology officer at Charge Robotics, used GPT-5.6 Sol through the Daybreak Blue program to discover a severe vulnerability affecting Ubuntu Linux systems in two hours, subsequently reporting the issue to the Debian Project for patching.

OpenAI said it has added checks to block high-risk activities, such as automated attacks against critical infrastructure including power grids and water treatment plants, backed by a $1 billion initiative to support defensive security professionals protecting essential services.