The most striking number in OpenAI's GPT-6 Cyber release is not a benchmark score. It is 948: the count of Linux kernel vulnerabilities for which the model produced a working exploit that grants root on the machine.
OpenAI published the model on 5 November and attached the word it uses sparingly — Critical — under its Preparedness Framework, the internal tier system for dangerous capabilities. It is the first OpenAI model rated Critical for cybersecurity; its predecessor, GPT-5.6 Cyber, sat one level below at High. Critical is the top of the scale.
In the Linux research, the model reproduced 1,724 vulnerabilities and turned 948 of them into working local privilege-escalation exploits, against 166 for GPT-5.6 Sol in the same setup. OpenAI reported 304 critical findings to the kernel maintainers along with draft patches. The campaign covered roughly 7% of the kernel and cost about $1.2 million in compute — around $1,280 per exploit.
In Firefox, researchers used the model to find 81 previously unknown high-severity flaws. Mozilla's advisory for Firefox 159, published on 27 October, credits OpenAI for 41 of the 167 vulnerabilities fixed in that release, all rated high. On ExploitGym, which asks a model to exploit a given flaw on a remote system, GPT-6 Cyber completed 48.5% of tasks against 32.1% for its predecessor.
The model is not in ChatGPT. It is available through the API and Codex only to approved customers in the Daybreak Red program, the tier reserved for authorised offensive security work; OpenAI says it will widen access over time. By default it requires IP allowlists, monitors credentials for signs of theft, and says it blocks help with attacks on power plants, water treatment facilities and industrial control systems.
The question that decides how much this matters is whether the model stays inside its assignment. OpenAI says it "avoided exploiting out-of-scope services in 99.93% of tests without its deployment prompt." It declined to tell The Register what happened in the remaining cases. That gap carries weight, because the entire pitch is that a model this capable can be trusted to stay within the lines.
The release also lands in an uncomfortable week for AI oversight. On 8 November the Associated Press reported that staff at the federal centre that tests frontier AI had been asked to report informally to a White House office, and that the centre has had no permanent director since July. Voluntary standards, a vendor's own risk rating and a government evaluator with a $10 million budget: the guardrails around a model that autonomously writes root exploits are, for now, mostly the vendor's own.
For defenders the practical readout is less dramatic. Kernel and Firefox updates now carry fixes for flaws found by a machine at a pace human audit never matched. Install them. The next wave of AI-found CVEs is not coming; it is already in the queue.

