OpenAI launches GPT-6 Cyber, a model built to find and fix security flaws

OpenAI has introduced GPT-6 Cyber, a version of its GPT-6 model built for cybersecurity tasks. According to the company, the system is designed to find exploitable vulnerabilities, meaning security flaws that could actually be used to break into a system, rather than theoretical weaknesses. OpenAI says the model can also develop fixes for the problems it finds.

The company says GPT-6 Cyber is trained to stay within authorized scope. In security work, that term usually refers to the boundaries set by a client or organization before any testing begins, limiting which systems, networks or applications a tester (human or automated) is allowed to touch. OpenAI’s framing suggests the model is meant to operate inside that kind of sanctioned engagement rather than scanning or probing systems without permission.

Pairing vulnerability discovery with fix development in one system is notable on its face. Many existing security tools specialize in one half of that work: scanners and fuzzers that hunt for flaws, or separate tooling that helps engineers patch them once found. A model that OpenAI says can do both would shift some of that handoff work onto the AI system itself, at least for the cases it can handle.

For security practitioners, the immediate questions are practical ones that OpenAI’s announcement does not answer. It is not stated how GPT-6 Cyber enforces the authorized-scope boundary technically, what happens if the model encounters a system or target outside the scope it was given, or what kind of oversight is built into its workflow. It is also not stated which types of vulnerabilities the model is most effective against, how its performance compares to existing automated scanning or penetration-testing tools, or what track record it has on real-world codebases.

The source material provided for this report is limited to OpenAI’s own short description of the model: that it exists, that it is meant to find exploitable vulnerabilities and develop fixes, and that it is trained to stay within authorized scope. Details such as a release date, pricing, availability, benchmark results, or documentation of how the authorization controls work are not included in that description and are not yet verified.

Because GPT-6 Cyber is pitched explicitly as a tool for authorized use, security teams evaluating it will likely want to see how OpenAI defines and technically enforces that authorization before deciding how much autonomy to give the model in their own environments. Until more technical detail is available, how GPT-6 Cyber performs against existing vulnerability-research tools, and how reliably it respects the scope it is given, remains an open question in OpenAI’s own terms.