OpenAI released a model that only focuses on network security, called GPT-6 Cyber, which focuses on two things: finding vulnerabilities and writing exploit programs.
The report card is available. Helped researchers find 81 previously unknown high-risk vulnerabilities in Firefox, each with evidence and reproduction steps and handed them over to Mozilla, which has successively patched them.
The Linux kernel is even more exaggerated. It reproduced 1,724 known vulnerabilities and wrote working exploits for 948 of them. Under the same environment, the previous generation GPT-5.6 Sol only succeeded 166 times.
Running points were also given. ExploitGym tests whether you can use specified vulnerabilities to hit remote targets. GPT-6 Cyber 48.5% and GPT-5.6 Cyber 32.1%.
The capabilities have increased so much that OpenAI itself became nervous first.
It raised the network security risk rating of this model from High to Critical, and added special monitoring, credential protection and sandbox isolation, and additional tightening of operations involving power, water, and industrial control. There is a key figure: without adding any prompts, the model has a 99.93% probability of not touching services outside the task range. After adding prompts, the probability is 100%.
The concept of placing the remaining 0.07% in security research is a matter of opinion. Exploiting vulnerabilities is troublesome enough if you get out of the circle once.
So it’s not all open now. The first batch will only be available to approved Daybreak Red customers through API and Codex, and will be released gradually later.
Good news for the defender, provided they get it before the attacker. OpenAI has the final say on this matter, and outsiders can only see how it reviews it.