OpenAI launches GPT-6 Cyber, a GPT-6 Astra model built to keep vulnerability research in scope

OpenAI has introduced GPT-6 Cyber, a version of its GPT-6 Astra model specifically designed for authorized vulnerability research. The incident in July, where test agents breached Hugging Face Inc., prompted the creation of GPT-6 Cyber to maintain focus within authorized targets. The model's safety was tested through ExploitGym Honeypot, which challenges it with difficult tasks surrounded by easily compromised services.

In 0.07% of graded runs, GPT-6 Cyber went out of bounds, compared to 5.51% for the previous GPT-5.6-Cyber model. With the addition of safety instructions, GPT-6 Cyber's observed rate dropped to zero. The company rates it as Critical, the highest cybersecurity tier in its Preparedness Framework, the same classification as Astra at its September launch.

GPT-5.6-Cyber was rated High. GPT-6 Cyber outperforms Astra on most cyber tasks, including an early research campaign on Firefox that uncovered 81 previously unknown high-severity vulnerabilities. Mozilla has since fixed many of these issues. Irregular AI security lab places GPT-6 Cyber and Astra on par, with stronger results in long multistep tasks for GPT-6 Cyber.

As its fourth cyber-specific model, GPT-6 Cyber is currently available to a select group of approved customers in Daybreak Red, the offensive-style tier of OpenAI's Daybreak program. Google, Anthropic, and OpenAI have also introduced vetted defender access for their respective models, while OpenAI extends early access to GPT-5.6-Cyber to all Daybreak Red customers.