OpenAI unveils GPT-6 Cyber: rated “Critical” and open only to approved customers

OpenAI has introduced GPT-6 Cyber, a model built to hunt for vulnerabilities in software and to write the exploits that prove they exist. The announcement text I consulted carries no date; CyberScoop supplies it: 5 November 2026. According to the company, in this initial phase the model is available through the API and Codex only to “a select group of approved customers” of Daybreak Red, the program dedicated to in-depth exploit development and authorized security testing. Alongside it sits Daybreak Blue, which OpenAI describes as designed for verified defenders: people validating flaws and fixes should run into fewer unnecessary refusals from the model. Pricing and eligibility criteria for Daybreak Red are not given. OpenAI places GPT-6 Cyber at the “Critical” cybersecurity level in its own Preparedness Framework, the top rung of the scale. Its predecessor, GPT-5.6 Cyber, was rated “High”. BleepingComputer confirms the classification.

Now the numbers, all supplied by OpenAI. For Firefox, the company reports 81 new vulnerabilities that Mozilla rated high severity, plus 30 already-known issues that turned out to be more serious than originally reported. BleepingComputer writes that Mozilla has already fixed many of them and points to advisory MFSA 2026-118, covering Firefox 159 and dated 27 October 2026. In the part I reviewed, that document lists more high-severity entries than the 81 cited. Several are credited to reporters linked to OpenAI, including the entry “OpenAI Codex Security”, but none names GPT-6 Cyber. So the match with the company's figure is not proven. On the Linux kernel, according to OpenAI, the model reproduced 1,724 vulnerabilities and developed working exploits for 948 of them. GPT-5.6 Sol, using the same harness, had stopped at 166. The company also says it reported 304 critical findings to maintainers, each with a candidate patch. BleepingComputer specifies that the 948 exploits are local privilege escalations: they let an unprivileged user obtain root permissions. The same outlet also reports 2,308 potential flaws found in the kernel, a figure that OpenAI's announcement, in the part I consulted, does not confirm. I could not directly check the technical post on the Linux data, “A thousand fewer paths to root”, because the page returned an error. Then there is Brotli, the compression library used by Node.js and other software. OpenAI says it found a critical vulnerability there and validated exploits showing possible remote code execution in applications that embed it. Finally, on the ExploitGym benchmark, GPT-6 Cyber exploits 48.5% of vulnerabilities, against 32.1% for GPT-5.6 Cyber.

These results have not been independently verified, and CyberScoop describes them as internal tests. Neither Mozilla nor the kernel maintainers have commented. Even the status of the fixes depends on the source: according to CyberScoop, all the reported flaws have been fixed; according to BleepingComputer, Mozilla has fixed many of them.

On staying within the authorized scope, OpenAI reports the following figures, which also come from the company's internal tests. Without the safety prompts intended for deployment, the model avoided attacking out-of-scope services in 99.93% of cases, against 94.49% for the previous model. With the prompts, the share rises to 100%. OpenAI researcher Eric Wallace told CyberScoop the model is “really good” at following instructions about what to touch and what to leave alone. The company lists several safeguards: misalignment monitoring specific to the cyber domain, targeted surveillance of emerging threats, and Private Safety Processing. On top of these come an IP allowlist enabled by default on the Daybreak Red API, checks for compromised credentials, and blocks on high-risk activity against critical infrastructure. OpenAI also recommends sandboxes, action review and human oversight. CyberScoop adds further identity and account-security requirements, as well as a partnership with Dragos.

The access model is the same one already seen with earlier cyber models from the big labs, OpenAI included. Offensive capabilities are deemed too risky for public release and remain reserved for verified customers. OpenAI's argument is that tools like this mostly serve defenders, provided the flaws get closed before someone else finds them. In the announcement, the company writes that it faces a “limited window” to use frontier models to secure software. It also says it wants to extend Daybreak to organizations of every size. BleepingComputer mentions independent researchers and open-source maintainers among the possible recipients.

What strikes me is that OpenAI itself puts its own model on the top rung of its risk scale and talks about a limited window. I read that as a warning sign more than a reassurance. What is still missing is the voice of those receiving the patches: until Mozilla and the kernel maintainers say how many of those flaws really bear GPT-6 Cyber's signature, the numbers remain a self-declaration, however detailed. — Olya