GPT-6 Cyber: OpenAI rates its new security model as Critical

OpenAI introduced GPT-6 Cyber on November 5. The model is trained to find exploitable flaws in software, confirm them with a working attack and stay within the limits that security researchers set for it. OpenAI's Preparedness Framework, its internal system for rating dangerous capabilities, puts GPT-6 Cyber at Critical for cybersecurity. Its predecessor, GPT-5.6 Cyber, was rated High. The Register notes that Critical is the framework's highest risk tier.

Access is limited to vetted security teams

The model is not part of ChatGPT. GPT-6 Cyber is initially available through the API and Codex, and only to a select group of approved customers in OpenAI's Daybreak Red program. That tier is meant for offensive security work, such as authorized penetration testing. OpenAI plans to open access to all Daybreak Red customers later. It requires IP allowlists by default, monitors credentials for signs of theft and says it blocks help with attacks on power plants, water treatment facilities and industrial control systems.

The benchmark results show why access is limited. On ExploitGym, which asks a model to exploit a given flaw on a remote system, GPT-6 Cyber completes 48.5% of tasks, compared with 32.1% for its predecessor. OpenAI also tested whether the model stays within its assignment. According to OpenAI, it "avoided exploiting out-of-scope services in 99.93% of tests without its deployment prompt." OpenAI declined to tell The Register what happened in the remaining cases. OpenAI models have gone beyond their instructions before, as our report on the model that broke into a government website during training shows.

What this means for Firefox and Linux users

OpenAI says its researchers used GPT-6 Cyber to find 81 previously unknown high-severity flaws in Firefox. Some have already been fixed. In its security advisory for Firefox 159, published on October 27, Mozilla credits OpenAI for 41 of the 167 fixed vulnerabilities, all rated high. We covered that release in our report on Firefox 159. The current version is Firefox 159.0.1, and the extended support releases are at 153.6 and 115.44. Anyone who has turned off automatic updates should check the version under Help and About Firefox.

In the Linux kernel, the model reproduced 1,724 vulnerabilities and built working exploits for 948 of them, giving an attacker on the machine root access. OpenAI reported 304 critical findings to the kernel maintainers along with draft patches, and says work on further fixes is ongoing. Distributions such as Ubuntu and Fedora will take time to ship those fixes, so it makes sense to install kernel updates promptly over the coming weeks. TechCrunch asks whether a model limited to select teams is enough, as AI labs face growing scrutiny over how they protect digital infrastructure. Our feature on self-copying AI targeting cheap TV boxes looks at what similar tools could do in the wrong hands.