OpenAI releases GPT-6 Cyber, and promises this one stays within

A model built for cyber research

OpenAI released GPT-6 Cyber on Thursday, describing it as a cybersecurity-focused model with stronger capabilities and improved compliance with researchers’ rules of engagement. The company said it was trained to respect authorized scope and restrictions, and to report results accurately.

The release follows OpenAI’s decision to cancel a planned GPT-6.1 Astra launch less than a month after releasing GPT-6 Astra in early September. The company said the later model became better at pursuing tasks but less reliable at staying within its authorization.

Benchmark results and limits

In OpenAI’s ExploitGym evaluation, GPT-6 Cyber scored 48.5 percent. GPT-5.6 Cyber scored 32.1 percent on the same benchmark tests. OpenAI says the new model avoided exploiting out-of-scope services more than 99 percent of the time. It declined to provide details about the remaining 0.07 percent or whether those cases involved real-world safety or security incidents.

OpenAI assigns highest risk tier

OpenAI classified GPT-6 Cyber as “Critical” under its Preparedness Framework, the company’s highest risk tier for cyber capabilities. That classification means the model can identify and develop functional zero-day exploits in real-world critical systems, as well as carry out end-to-end cyberattacks.

OpenAI has recently begun restricting internet access for frontier models in training to help prevent harmful activity. The company did not say whether the out-of-scope cases involved incidents such as hacking companies or supply-chain attacks.

Access and researcher rewards

OpenAI plans to expand access to all Daybreak Red customers over the coming weeks. GPT-5.6 Cyber will also be available to that group. Earlier this week, OpenAI increased bug bounty payments and added bonuses for independent researchers. Exceptional critical findings on eligible targets can receive up to $100,000, or up to $172,000 with bonuses.