OpenAI has released GPT-6 Cyber, a version of its model built for offensive and defensive security work. The company says the model operates within so-called rules of engagement, the limits set for each task. The report was published by The Register on November 5, 2026.
On ExploitGym, a benchmark the company uses to measure how well a model exploits vulnerabilities, GPT-6 Cyber reached 48.5%. The previous version, GPT-5.6 Cyber, scored 32.1% on the same tests. In a separate evaluation that checks whether the model respects the scope of an assigned task, OpenAI says it stopped exploiting out-of-scope services in more than 99% of cases. Only 0.07% of cases went past the boundary set for the exercise.
The launch is part of OpenAI's push to expand its footprint in cybersecurity, a field where the ability to find and exploit vulnerabilities is exactly what makes a tool valuable and dangerous at once. The figures come from the company itself, with no published external review so far, and the gap between benchmark performance and behavior in real environments has not yet been demonstrated.
For security teams running production systems, the key point is that the promise of containment depends on configuration and on controls outside the model. A 99% scope-adherence rate still leaves room for improper actions during long operations. Access to tools like this should therefore be limited to isolated environments, with every action logged and explicit authorization for each target. Trampolim's view is that the launch expands defensive and offensive capability at the same time, and that internal governance needs to keep pace with that change.