Picture a security researcher at two in the morning, scrolling through millions of lines of Linux kernel code. Somewhere in that endless text hides a flaw that could hand a stranger the keys to the machine.
Finding it by hand could take weeks. OpenAI says its newest model did that job — 948 times over — in a single automated run.
On November 5, OpenAI introduced GPT-6 Cyber, a specialised version of its flagship GPT-6 Astra model built for one purpose: hunting software vulnerabilities, proving they can be exploited, and helping developers fix them.
The numbers OpenAI attached to the launch are startling. They also explain why the company gave the model its highest possible risk rating.
According to OpenAI, GPT-6 Cyber reproduced 1,724 vulnerabilities in the Linux kernel and turned 948 of them into working local privilege escalation exploits — attacks that let an unprivileged user gain root access.
Mozilla, meanwhile, classified 81 previously unknown Firefox flaws found by the model as high severity. Both campaigns ran in controlled research settings, and the findings were disclosed to the maintainers before anything was published.
What GPT-6 Cyber Actually Does
GPT-6 Cyber is not a general-purpose chatbot with a security hobby. It is a version of GPT-6 Astra trained specifically for authorised security work: reading source code, spotting weaknesses, verifying that a weakness is genuinely exploitable, and drafting a fix.
The difference matters because finding a bug and proving it matter are two very different jobs. A scanner can flag thousands of suspicious lines. A working exploit proves that one of them actually opens a door.
OpenAI built GPT-6 Cyber to do the second job — the one that normally takes a skilled human hours or days per flaw.
The model is also designed to stay inside the boundaries researchers set for it. In OpenAI’s testing, GPT-6 Cyber avoided out-of-bounds services 99.93 percent of the time, up from 94.49 percent for its predecessor, GPT-5.6 Cyber.
That reliability, the company says, is what makes the model usable in real security programmes rather than just in a lab.
948 Linux Exploits in a Single Automated Run
The Linux kernel campaign is the headline. OpenAI ran GPT-6 Cyber against the latest Linux kernel development code using an automated pipeline that searched for candidate flaws, reproduced them, and attempted to turn each one into a working exploit.
The pipeline flagged 2,308 potential flaws. It reproduced 1,724 of them. And it produced working local privilege escalation exploits for 948 distinct vulnerabilities — more than five times the 166 that GPT-5.6 Sol managed with the same research setup. According to Reuters and Fortune, 588 of those 948 flaws were previously unknown.
The run was not cheap. OpenAI says the research consumed about $1.21 million in API credits, working out to roughly $1,279 per working exploit. That sounds expensive until you compare it with the alternative: months of human researcher time for the same coverage.
OpenAI says it has passed 304 critical findings to Linux maintainers, along with proposed patches, following the community’s own guidance for disclosure. The goal, the company insists, is a thousand fewer paths to root — the title of its research post — not a thousand new weapons.
81 Firefox Flaws Mozilla Did Not Know About
The second campaign targeted Firefox, one of the most scrutinised codebases on the planet. GPT-6 Cyber produced evidence and reproducible tests for 81 previously unknown vulnerabilities that Mozilla classified as high severity.
It also found that 30 known issues were worse than first reported. OpenAI shared the findings with Mozilla along with proposed patches, and Mozilla began landing fixes shortly after receiving the reports.
Firefox engineer Brian Grinstead put the result in perspective. The browser entered 2026 as one of the world’s most hardened codebases, he said, yet the model still found new high-severity bugs. That is the point that should worry every software team: if a model can find flaws in Firefox, it can find them almost anywhere.
Mozilla has separately confirmed that it works with frontier models from OpenAI and Anthropic to speed up Firefox security fixes, so the collaboration behind the numbers is real and ongoing.
Why GPT-6 Cyber Scored 48.5 Percent on ExploitGym
Benchmarks tell the same story of a sharp jump. On ExploitGym, a benchmark that tests whether a model can exploit a specified vulnerability on a remote system, GPT-6 Cyber scored 48.5 percent. Its predecessor, GPT-5.6 Cyber, scored 32.1 percent. Even the general-purpose GPT-6 Astra managed only 42.4 percent.
The gap between the specialist and the generalist is the whole argument for GPT-6 Cyber’s existence. A model trained for security work, with the right tools and the right guardrails, outperforms a stronger general model at the security task. Specialisation, in this case, beats raw scale.
OpenAI is careful to note that every score and count it published is its own claim unless another source is named. There is no independent verification yet of every finding the model produced, and the full deployment safeguards have not been disclosed in detail.
The numbers are impressive, but they come from the company that built the model.
OpenAI’s First “Critical” Cyber Rating
Here is the twist that makes GPT-6 Cyber different from every model OpenAI has released before. Under the company’s Preparedness Framework, GPT-6 Cyber is classified as “Critical” for cybersecurity — the highest designation on the scale. GPT-5.6 Cyber was rated “High”.
The rating reflects a dual-use problem that OpenAI does not hide. A model that can find vulnerabilities and write working exploits can defend software, but the same capability in the wrong hands can attack it. The company says it has added safeguards for industrial controls, essential services, election systems and other high-risk targets.
That is why access is restricted. GPT-6 Cyber is available only through the API and Codex to a select group of approved customers in Daybreak Red, OpenAI’s programme tier for advanced security work such as exploit development and penetration testing. Access will expand over time, OpenAI says, but not to everyone, and not soon.
Who Gets to Use GPT-6 Cyber
The early customer list reads like a security industry roll call. OpenAI named Cisco, Calif, CodeRabbit and Palo Alto Networks among the teams already using the model for vulnerability research and code review, alongside its own security teams and Mozilla’s.
The price of entry is steep in more than money. Organisations need single sign-on, multi-factor authentication and IP allowlisting. Individuals need identity and liveness checks plus a physical security key. API pricing is set at $20 per million input tokens and $100 per million output tokens.
For the teams that qualify, the pitch is simple: a tireless researcher that never sleeps, never gets bored reading kernel code, and can test a thousand hypotheses before lunch. For everyone else, GPT-6 Cyber is a reminder that the most powerful security tools are now also the most closely guarded.
What Happens Next for AI Security
GPT-6 Cyber arrives in a week when AI security is moving fast.
Microsoft’s own researchers recently showed that reprogramming fewer than ten dormant neurons could backdoor an AI agent, a finding that underlines how fragile even well-tested systems can be. OpenAI’s answer is a model that hunts flaws before attackers do — but only in approved hands.
The company has been building safety infrastructure around its models all year, from the Control Room kill switch for rogue AI models to this new hunting tool.
The stakes were underlined by the October wave of AI infrastructure breaches, including the Fireworks AI breach that exposed customer datasets and model weights.
Every vulnerability GPT-6 Cyber finds and discloses makes software safer. But the same techniques, in a model with fewer guardrails, would make attackers faster too. OpenAI’s bet is that controlled access, identity checks and responsible disclosure keep the advantage on the defenders’ side.
OpenAI has a security keynote scheduled for Thursday, November 12, where more detail on its security roadmap is expected.
Until then, the numbers speak for themselves: 948 Linux exploits, 81 Firefox flaws, and the first “Critical” rating OpenAI has ever handed to one of its own models. The age of AI-powered vulnerability hunting has arrived — and, for now, only a chosen few get to use it.
Source: OpenAI’s GPT-6 Cyber announcement, as reported by MLQ.
