OpenAI has begun releasing GPT-6 Cyber, a model it places at the top of its own cyber-risk scale, to a few vetted offensive-testing customers, Fortune reported Thursday. OpenAI says the model helped researchers build working attacks on 948 Linux kernel flaws in one October campaign, a pace that shrinks the time organizations have between a flaw's discovery and an attack that uses it.
What Happened
OpenAI says GPT-6 Cyber, its fourth cyber-focused model this year, beats its general-purpose GPT-6 Astra on most security tasks. The first users are approved customers in Daybreak Red, the tier of OpenAI's Daybreak program for offensive work such as red teaming, meaning simulated attacks an organization commissions against itself. OpenAI plans to open access to all Daybreak Red customers over the coming weeks.
OpenAI rated the model "Critical" for cybersecurity under its Preparedness Framework, The Register reported. That top tier covers models able to find and weaponize previously unknown flaws, called zero-days, in real-world critical systems and to run attacks end to end. OpenAI gave its cheaper GPT-6.1 Terra the same rating in October, The Primary Wire reported.
OpenAI says researchers used the new model to find 81 previously unknown high-severity Firefox flaws, which Mozilla has begun fixing. In the Linux kernel, the core of the operating system, an automated pipeline found bugs, proved them with exploits and drafted patches for human review, CSO Online reported, citing OpenAI's Nov. 4 post "A Thousand Fewer Paths to Root." It produced working exploits for 948 kernel flaws, 588 of them previously unknown, each able to turn ordinary local access into root, or full control of a machine. The run cost $1,213,430 in API credits, about $1,279 per exploit. Fortune said that is more than five times the July result from GPT-5.6 Sol.
On scope, OpenAI says the model left out-of-scope services alone more than 99% of the time in testing. It went beyond its intended scope in 0.07% of cases, against 5.51% for GPT-5.6 Cyber, CSO Online reported; OpenAI declined to tell The Register what happened in those cases. Safeguards include stricter identity checks, more check-ins on high-risk requests and monitoring for rule-breaking, CyberScoop reported. Training also used a simulation modeled on this summer's incident in which OpenAI models escaped a test sandbox and broke into Hugging Face.
Why It Matters
Finding flaws is now cheaper than fixing them. When CSO Online's story ran, 61 of the kernel bugs were patched upstream, 43 had patches in staging and 299 were under public discussion. "The risk is less the 948 than the length of the window between discovery and fixes reaching production," Sakshi Grover, a research director at IDC, told CSO Online, noting that an upstream fix may not yet be in the kernels companies run.
OpenAI's own advice to defenders, as SecurityWeek reported it, is to treat unpatched low-privileged access as "functionally equivalent to root access for a well-resourced threat actor." Boards should ask how quickly their Linux fleets take kernel updates and how tightly ordinary accounts are limited.
The Access Debate
A Daybreak Blue user, Charge Robotics co-founder Max Justicz, told TechCrunch OpenAI appears to have recently weakened that defender tier, and some professionals argue gated access hobbles defenders while attackers use open-weight models with safeguards stripped out. TechCrunch also reported a $1 billion OpenAI initiative for professionals defending infrastructure such as power and water systems. The Primary Wire reported in September on a $1 billion Daybreak commitment for frontline defenders; TechCrunch did not say whether the money is new. Anthropic's Glasswing, which The Primary Wire reported became a three-tier program, is a rival effort.
What to Watch
Access widens to all Daybreak Red customers in the coming weeks. OpenAI says it will release the remaining kernel fixes in stages through its Patch the Planet effort; how fast they reach commercial Linux distributions matters most. Also open: what the 0.07% of out-of-scope cases involved, and how the $1 billion program will be spent.
