OpenAI Limits GPT-6 Cyber to Approved Security Teams

OpenAI released GPT-6 Cyber on November 5, 2026, a security model built to find software flaws, prove they can be exploited, and help developers fix them. The company is limiting first access to a small group of approved customers in its Daybreak Red programme, which covers authorised offensive security work such as penetration testing.

The model is a version of the company's flagship system trained for authorised vulnerability research. According to OpenAI, it outperforms that flagship system on most cybersecurity tasks, while receiving extra training to stay inside the scope a researcher sets. That boundary matters because the same skills that find flaws in a test environment could cause harm outside it. It is the fourth cyber-focused model OpenAI has released this year, following versions in April, June and August.

What the model found in Firefox and Linux

In a campaign against Firefox, the model produced reproducible evidence for 81 previously unknown vulnerabilities that Mozilla classified as high severity, according to OpenAI. It also showed that 30 known issues were more serious than first reported. Mozilla has released fixes for many of the flaws, the company said. Security teams at Mozilla, Palo Alto Networks, Cisco and CodeRabbit are among those using the model for vulnerability research and code review.

A separate campaign against the Linux kernel reproduced 1,724 vulnerabilities and turned 948 of them into working local privilege escalation exploits, which take an ordinary local user to root access. The same research setup produced 166 working exploits with an earlier model. OpenAI said it disclosed 304 critical findings to software maintainers with candidate patches and is working through further reports with partners in its Patch the Planet programme.

The release follows a research post published the day before, titled "A thousand fewer paths to root", which described an autonomous pipeline that searches for candidate flaws, reproduces them and attempts to turn them into working exploits. OpenAI said the pipeline ran against the latest Linux kernel development code. The company also said its researchers used the model in controlled campaigns against its own infrastructure, where agents chained vulnerabilities to escape microVMs and reach underlying hosts, prompting patches and tighter network restrictions.

Stronger scores, tighter limits

On ExploitGym, a benchmark that tests whether a model can exploit a specified vulnerability on a remote target, the new system scored 48.5%. Its predecessor scored 32.1%, a gap OpenAI describes as a 51% relative improvement, while the company's flagship model scored 42.4% on the same test. On a new benchmark called RemoteExploitBench, the model completed 95% of exploit development tasks.

OpenAI also tested whether the model stays inside authorised scope. In a test called ExploitGym Honeypot, which leaves out-of-scope services deliberately easy to attack, it avoided exploiting them in 99.93% of cases without a deployment safety prompt, compared with 94.49% for its predecessor, and in every case with the prompt in place. The company classifies the model as Critical for cybersecurity under its Preparedness Framework, the highest tier, up from High for the earlier system.

Who gets access and why it is restricted

Access runs through the API and Codex, and requests must come from approved networks by default. OpenAI pairs the release with identity checks, organisational security requirements, monitoring, sandboxing, credential protections and human oversight. The company said the restrictions reflect the risk that the same capabilities could support abuse, so the model is not being offered as an unrestricted public release. The launch comes during a volatile week for AI shares, after an AI stock selloff erased the post-midterm rally in a single day.

The earlier model is now available to all Daybreak Red customers, and OpenAI is widening access to its Daybreak Blue programme, which serves verified defenders. A separate track covers approved critical infrastructure defenders, and the company is subsidising access for resource-constrained defenders through a grant programme worth one billion dollars.

Greg Brockman, an OpenAI co-founder, described the release in a post on X as "a major advance for security research, already being used to help defend security-critical software such as the Linux kernel, Firefox, and many other projects." Rival lab Anthropic also featured in AI infrastructure news this week. For now, the tool stays in the hands of a small group of vetted teams. OpenAI said access will widen to all Daybreak Red customers over time. Sources: OpenAI announcement and Neowin report.