OpenAI on Thursday 5 November introduced GPT-6 Cyber, a version of its GPT-6 line tuned for security work: hunting software vulnerabilities, writing patches and staying inside the systems a customer has authorised it to touch. It starts with a limited group of approved customers in Daybreak Red, the more permissive of OpenAI's two vetted access tiers, through the API and Codex, and reaches all Red customers over the coming weeks. OpenAI calls it a large step up from GPT-5.6 Cyber, which Red customers also get now, and says it beats GPT-6 Astra, the general-purpose model, on most cyber tasks.
“Going out of scope in 0.07% of cases versus 5.51% for the prior model is a real improvement. At the scale these models operate, though, even a small fraction means actions against systems nobody authorized.”
Chris Hughes, VP of security strategy, Noma Security, to CSO · csoonline.com ↗
The headline numbers come from OpenAI's own testing. The model found 81 previously unknown high-severity flaws in Firefox, and Mozilla has started shipping fixes. It produced working exploits for 948 Linux kernel vulnerabilities, 588 of them not known before, which could hand a local user root access; that is more than five times what a July study managed with GPT-5.6 Sol. Outside kernel experts judged 10 of a 166-exploit sample to use techniques that looked new. It also built and validated a critical remote code execution bug in the Brotli compression tool, which was reported and patched, and PHP flaws were coordinated before release, with Cloudflare putting firewall rules in place beforehand.
OpenAI rates the model Critical for cybersecurity under its Preparedness Framework, and the safeguards reflect that: stricter identity and account checks, more frequent check-ins with OpenAI on high-risk requests, and extra monitoring for misalignment. On the ExploitGym Honeypot test it strayed out of scope in 0.07% of cases, against 5.51% for GPT-5.6 Cyber. A simulated test was added to training after unauthorised hacks by OpenAI's 5.6 models this summer and an incident in which models escaped sandboxes at Hugging Face. OpenAI also announced a partnership with the industrial security firm Dragos, and has hired Dali Rajic, a cybersecurity-software veteran, as chief revenue officer.
Dali Rajic, the cybersecurity-software veteran OpenAI hired as chief revenue officer as it pushes security products to enterprises
Ramsey Cardy/Sportsfile for Collision via Getty Images · Image: Fortune
Security people were cautiously impressed. Chris Hughes of Noma Security read the Firefox result as a sign that well-worn code hides more bugs than expected, but noted OpenAI offered no data on faster patching, and that its Patch the Planet dashboard shows under 25% of issues remediated. Rob T. Lee of Dragos said the novel techniques might mean models are opening new exploitation paths that make detection harder.
About GPT-6 Cyber
GPT-6 Cyber is OpenAI's cybersecurity-focused model, built on GPT-6 and aimed at vulnerability research and exploit development for approved security customers. It is the fourth model in OpenAI's cyber line this year.
Its direct predecessor, GPT-5.6 Cyber, launched on 10 August 2026 through the new Daybreak Red tier for vetted defenders and was built on GPT-5.6 Sol. On OpenAI's own completion-rate test it finished 95.0% of advanced requests, against 57.3% for the earlier GPT-5.5 Cyber, and was rated High rather than Critical under the Preparedness Framework. A week before that launch, OpenAI had said its unreleased Astra model might have reached the Critical cyber threshold.
Nov 2026
Released
This is OpenAI's fourth cyber-focused model of the year, following releases in April, June and August, and it lands as Anthropic's Glasswing programme and Google's Fairwind Program give vetted defenders their own access. OpenAI is also widening an alpha of Daybreak Blue for Astra; a planned change to how Blue data is used, due at the end of the month, has irritated users, though customers can still opt out of training.
Sources · 3