Back to the stories

OpenAI releases GPT-6 Cyber with restricted initial access for cybersecurity agents

Score 10.0,

Artificial intelligence

AI that can find a software weakness is useful only if it knows where it is allowed to look.

That is the challenge behind GPT-6 Cyber, OpenAI’s new model for cybersecurity work. It is designed to find and confirm vulnerabilities, help develop fixes, and stay within boundaries set by researchers. Access is starting with a select group of approved Daybreak Red customers.

The promise is practical: security teams could find and fix flaws sooner. But the same ability to uncover weaknesses could cause harm if a model acts beyond the systems it was asked to test.

OpenAI says GPT-6 Cyber performed 51% better than its previous cyber model on a vulnerability-finding test. In a separate isolated test, the company says it stayed within authorized boundaries every time when given deployment safety instructions.

Think of a security tester given permission to inspect one building. The job is to find weak doors there, not try the doors next door. OpenAI says it uses safeguards such as sandboxing and human review. But results in controlled tests do not prove a model will never cross a line in real use.

For now, this is not a tool anyone can simply sign up to use. OpenAI is limiting access to approved customers, with checks around how the model operates. That makes the rollout a test of both capability and control.

The real question is whether those boundaries hold when security teams put the model to work on real systems.