OpenAI releases GPT-6 Cyber to approved customers and rates it Critical for cybersecurity
The figures come from OpenAI, and no independent rerun was found. Critical is OpenAI's own classification. This edition read the announcement, the Linux report and The Register, not the system card or Mozilla's advisories. The Register reports that OpenAI declined to detail the 0.07% of out-of-scope cases, including whether any involved a real-world incident. The Linux exploits need existing local access; they are not remote attacks, and they were made in a controlled research environment. Test results do not show how the model will behave in customer use.
What was claimed?
On November 5, 2026, OpenAI released GPT-6 Cyber, a model for finding and validating exploitable software flaws, first to a select group of approved Daybreak Red customers. OpenAI classifies it as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. It reports 48.5% on its ExploitGym test, up from 32.1%, and says the model avoided out-of-scope services in 99.93% of a separate test without its deployment prompt, up from 94.49%, and in 100% with that prompt. In a November 4 report, OpenAI said a pipeline using an upcoming cyber model examined 2,308 candidate Linux kernel findings, reproduced 1,724 vulnerabilities and produced 948 working local privilege escalation exploits at a cost of $1,213,430 in API credits. The November 5 post attributes the 948 exploits to GPT-6 Cyber and also reports 81 previously unknown high-severity Firefox vulnerabilities.
What does the evidence support?
Company claim, independent confirmation unresolved
The Register describes Critical as OpenAI's highest risk tier for cyber capabilities, meaning a model can find and build working exploits against real critical systems and carry out end-to-end attacks. If the figures hold, finding and proving serious flaws can run as a systematic process at about $1,279 per Linux exploit. That helps defenders, and it raises the stakes of misuse, which is why access is limited and why OpenAI stresses training the model to respect limits.
What are the limits?
The figures come from OpenAI, and no independent rerun was found. Critical is OpenAI's own classification. This edition read the announcement, the Linux report and The Register, not the system card or Mozilla's advisories. The Register reports that OpenAI declined to detail the 0.07% of out-of-scope cases, including whether any involved a real-world incident. The Linux exploits need existing local access; they are not remote attacks, and they were made in a controlled research environment. Test results do not show how the model will behave in customer use.
Claim or evidence date: 2026-11-05. Last verified: .
OpenAI · GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents ↗
Colors classify evidence; they do not rank truthfulness. Explore hype versus reality →