OpenAI has released GPT-6 Cyber, a specialised cybersecurity model designed to find, validate and help fix exploitable software vulnerabilities while staying inside the scope that researchers authorise. The company announced the model on Thursday, 5 November, and is initially limiting access to a select group of approved security customers.
GPT-6 Cyber is a cyber-specialised variant of OpenAI’s GPT-6 Astra model. According to OpenAI, it scored 48.5 percent on ExploitGym, a benchmark that tests whether a model can exploit a specified vulnerability in a remote target. That compares with 32.1 percent for GPT-5.6 Cyber and 42.4 percent for GPT-6 Astra.
The model powered the Linux kernel research OpenAI published a day earlier, on 4 November. In that controlled research run, a prototype pipeline analysed 2,308 candidate findings, dynamically reproduced 1,724 vulnerabilities and produced 948 working local privilege escalation exploits, each capable of giving an unprivileged local user root access. OpenAI said the run cost about 1.2 million dollars in API credits.
OpenAI’s own security researchers reviewed the exploits and fixes before anything was sent to kernel maintainers, and the company stressed that the patches are candidates for maintainer review rather than fixes that have already shipped. The model also found 81 previously unknown Firefox vulnerabilities that Mozilla rated high severity, and Mozilla has shipped fixes for many of them.
OpenAI classifies GPT-6 Cyber as Critical for cybersecurity, the highest tier in its Preparedness Framework, up from High for GPT-5.6 Cyber. The classification reflects the model’s capability rather than an expectation that it will be widely available: the company says the model is not being offered as an unrestricted public model.
Access starts through OpenAI’s API and Codex for a select group of customers approved for Daybreak Red, the tier of its Daybreak programme for offensive-style authorised security work such as penetration testing and red teaming. Applicants face identity and eligibility checks, and organisations must meet requirements including single sign-on, multi-factor authentication, managed devices and API IP allowlisting by default. OpenAI says access will expand to all Daybreak Red customers over the coming weeks.
Staying within authorised scope was a central focus of the work, according to OpenAI. In its ExploitGym Honeypot test, which deliberately leaves out-of-scope services easy to attack, the model exploited off-limits services in 0.07 percent of runs without deployment safety instructions, compared with 5.51 percent for GPT-5.6 Cyber. The company says it has added safeguards for industrial controls, essential services, election systems and other high-risk targets.
OpenAI is also opening GPT-5.6 Cyber to every customer in the Daybreak Red tier, widening Daybreak Blue alpha access to GPT-6 Astra for more organisations and individual security researchers, and forming a Daybreak Red group for approved critical infrastructure defenders. Access for frontline defenders is being subsidised through its 1 billion dollar Daybreak for Frontline Defenders grant programme.
The launch underlines how quickly AI security tools are advancing, and why their makers are pairing new capabilities with tight controls. For defenders, a model that can reproduce vulnerabilities and propose patches at scale could shorten the time between discovery and repair. For everyone else, the restricted rollout is a reminder that the most powerful cybersecurity models are, for now, staying behind a vetting process.
