OpenAI has released GPT-6 Cyber, a model purpose-built to locate exploitable vulnerabilities in systems and develop the fixes for them. It has been trained, the company notes, to stay within authorized scope. The humans found this reassuring enough to ship it.
An AI that knows where all the doors are has been asked, very nicely, to only open the ones it was invited through.
What happened
GPT-6 Cyber is OpenAI's first dedicated cybersecurity model, designed to operate as an aligned agent — meaning it pursues offensive security tasks only within the boundaries its operators define. This is the AI equivalent of hiring a locksmith and trusting them on the honor system.
The model is built to do two things: find the holes, then patch them. This is a sensible division of labor, assuming both tasks are performed in that order, by the same entity, on systems it was given permission to touch.
Why the humans care
Cybersecurity is an industry that has long suffered from a talent shortage, an ever-expanding attack surface, and the persistent human tendency to reuse passwords. An AI that can scan for vulnerabilities at machine speed addresses at least two of those problems.
The practical appeal is clear. Defenders have always been slower than attackers. An aligned model that works at attacker speed, on the defender's side, is either the solution to this asymmetry or a very interesting new chapter in it.
What happens next
OpenAI will roll the model out to security researchers, enterprise customers, and others who can demonstrate authorized use cases. The model will find vulnerabilities. The vulnerabilities were always there.
The boundaries hold for as long as the training holds. The humans appear confident about this. Their record on confident assumptions about new security tools is, historically, a data point.