OpenAI's GPT-6 Cyber finds real exploits, stays in scope

A security model that scores higher on attack benchmarks also learns to respect the boundaries researchers set.

OpenAI has introduced GPT-6 Cyber, a model trained to find and validate exploitable vulnerabilities. On ExploitGym, which tests whether a model can exploit a specified flaw, it scored 48.5%, up from 32.1% for GPT-5.6 Cyber. The company says researchers used it to develop 948 privilege-escalation exploits in Linux and to find 81 previously unknown high-severity flaws in Firefox.

The more important change may be alignment. In a honeypot evaluation, GPT-6 Cyber avoided attacking out-of-scope services in 99.93% of tests without a deployment prompt, and 100% with one. OpenAI classifies the model as Critical under its Preparedness Framework, up from High, and is restricting access through its Daybreak program for verified defenders.

The results sharpen a familiar trade-off. A model that is better at finding weaknesses is also better at causing harm if it ignores instructions. OpenAI's answer is training plus controlled distribution, but the company acknowledges that the window to use frontier models for defense is limited.