GPT-6 Cyber: 948 Ways to Root a Linux Box, US$1,279 Each
Petr Červenka
Listen to this article · 12 min
On this page 8 sections
The short version
- On 5 November OpenAI announced GPT-6 Cyber, a model built to find security holes, prove they are real and help write the fix. It is not public: only approved security customers get it.
- In one October run it built 948 working exploits that turn an ordinary Linux user into root (the all-powerful admin account). The run cost US$1.2 million, or about US$1,279 per exploit.
- It also found 81 previously unknown high-severity flaws in Firefox. Mozilla is shipping the fixes, starting with Firefox 159.
- OpenAI's own advice: a low-privilege account on a Linux machine that is behind on patches should be treated as root.
- For WA businesses, the to-do list is short and boring: patch Linux and reboot, keep browsers current, and stop giving untrusted code a shell on shared machines.
Linux's mascot is a penguin called Tux. Last month OpenAI found 948 new ways to make him do whatever you like.
OpenAI announced GPT-6 Cyber on , a day after OpenAI published the Linux research behind it. BleepingComputer and others picked it up the following day. This post explains what the model did in plain English, what it cost, and what it means if your business runs on Linux servers or Firefox, which is most businesses whether they know it or not.
One thing up front: every number below comes from OpenAI, which sells this model. The Firefox part we checked against Mozilla's own advisory. The Linux part we can't check independently yet, though the patches are landing in public.
What is GPT-6 Cyber?
It is a version of OpenAI's models trained for security research: finding weaknesses in software, proving they can be used, and drafting the patch. OpenAI rates it "Critical" for cyber capability under its own risk framework. That is the top of its scale, one step up from the "High" it gave the previous version, GPT-5.6 Cyber.
On ExploitGym, a test of whether a model can break into a system through a known flaw, it succeeded 48.5% of the time. GPT-5.6 Cyber managed 32.1%.
You can't sign up for it. It is available, through OpenAI's API and its Codex coding tool, only to a small group of approved customers in a program called Daybreak Red. OpenAI says it will widen that to all Daybreak Red customers soon, and wants to bring Daybreak to "organizations of every size, independent researchers, and open-source maintainers". Defenders with fewer needs can use Daybreak Blue, which runs OpenAI's mainline models with fewer refusals for verified security work.
The Linux run, by the numbers
In mid-October OpenAI pointed a pipeline of AI agents at the latest Linux kernel, the core of the operating system under most of the world's servers, Android phones and cloud platforms. Human researchers reviewed what came out at the end.
| Stage | What happened |
|---|---|
| Pick targets | "Scout" agents chose 5,034 parts of the kernel an ordinary user can reach: about 2.8 million lines, 7% of the code |
| Look for bugs | "Analyst" agents produced 2,308 candidate flaws in 12 hours |
| Prove them | In 30 hours, agents reproduced 1,724 of them (75%) on test machines |
| Break in | 948 distinct flaws got working exploits that take an ordinary user to root, plus 291 that leak memory |
| Fix | Candidate patches for each, reviewed by humans before going to the Linux maintainers |
The test machines ran a standard desktop Linux with the usual protections switched on. OpenAI did not weaken the target to make the numbers look good.
The receipt
The whole run cost US$1,213,430 in API credits, or US$1,279 per working root exploit. The exploit stage alone averaged US$359 per success.
For comparison, OpenAI notes that the Pwn2Own and TyphoonPWN hacking competitions have historically paid US$20,000 and US$70,000 for a single Linux root exploit. At those prices the 948 exploits would be "worth" US$18.9 million to US$66.3 million. Even before anyone has optimised it, the pipeline makes a profit.
It's getting better quickly
The same pipeline, run in July with the older GPT-5.6 Sol model, found 166 root exploits for US$513,690. Three months later the new model found 948, more than five times as many, in a kernel that had been hardened in between.
The agents also used at least 10 exploitation techniques that outside kernel experts had not seen before. OpenAI has not published the details, for obvious reasons.
The bit we liked most
In an early run, the pipeline reported 200 successful exploits. Human review found that 20 of them got root through a different unknown flaw from the one they had been told to use. The agent got the right answer the wrong way. OpenAI now checks that each exploit uses the flaw it was assigned.
If you run AI agents anywhere near real systems, the lesson is worth copying: check how the agent got the result, not just whether it got there.
What about Firefox?
OpenAI says GPT-6 Cyber found 81 previously unknown flaws that Mozilla rated high severity, and showed that 30 already known issues were worse than first reported. It sent Mozilla test cases and proposed patches. Brian Grinstead, a Distinguished Engineer on Firefox, is quoted in the announcement saying Firefox entered 2026 as one of the most scrutinised codebases around, and the model still found new high-severity bugs.
Mozilla's advisory for Firefox 159, released on 27 October, fixes 167 vulnerabilities, 108 of them high severity. OpenAI is credited as the reporter on 41 of them. OpenAI says Mozilla "has released fixes for many" of the 81, which suggests the rest are still to come.
If your staff use Firefox, make sure it is on version 159 or later. Firefox updates itself, but only applies the update when it restarts, and some people never close their browser.
Why this matters if you are not OpenAI
You might reasonably think: good, the people with the scary model are the ones fixing things. That's partly true. Linux maintainers moved fast. Ten days after the scan, 61 of the flaws had been patched upstream, 43 had patches staged, and 299 were being discussed in public. OpenAI says it has sent 304 critical findings to the maintainers, each with a candidate patch.
Here's the catch. A patch only protects you once you install it. Every public kernel fix shows an attacker where to look, and as GLM-5.3 showed in September, turning a published fix into a working attack is now cheap even with models anyone can download. OpenAI says so itself: within a few years, it expects even less capable models to be able to build exploits for flaws once they are disclosed.
OpenAI also put its practical advice plainly. For now, an ordinary user account on a Linux system that isn't fully patched should be treated as "functionally equivalent to root access" against a well-resourced attacker.
Think about where that matters:
- Shared servers and hosting. Every customer or app with a login on the same machine.
- CI/CD runners. Build machines that run code from pull requests, plugins and dependencies.
- Containers. They share the host's kernel. A kernel root exploit from inside a container is a way out of it.
- Anything with a web shell, file upload or plugin system where a bug could give an outsider a foothold as a low-privilege user.
This is the same pattern as this autumn's other big flaws. PHP2Shell and HTTP2Shell were both found with OpenAI models (OpenAI says GPT-6 Cyber helped prove the Brotli flaw could be used to run code remotely), and in both cases the race was between the patch and the attackers.
Does the Essential Eight still cover it?
The controls, yes. The timeframes are getting tight. Under the Essential Eight, internet-facing systems get 48 hours when a flaw is critical or already being exploited. Workstations and servers that don't face the internet can wait up to a month.
A month was a reasonable risk call when a kernel exploit took a specialist weeks to build. When it costs an AI about US$1,279 and nobody has had to be clever, an internal server with a dozen logins on it is a softer target than its "non-internet-facing" label suggests.
ASD's newer ASD Essentials asks for patches "within defined risk-based timeframes". This is the kind of news that should move your risk call.
What to do this month
- Patch the Linux kernel, then reboot. A kernel update does nothing until the machine restarts (or you use live patching). Check how long your servers have been up:
uptimemeasured in months means the kernel is months old. - Shorten the gap for shared machines. Hosting boxes, CI runners and container hosts should be patched like internet-facing systems, because the people with accounts on them may not be trustworthy.
- Don't treat a container as a wall. If you run code you didn't write, look at stronger isolation such as gVisor or Firecracker micro-VMs, which OpenAI also recommends.
- Update Firefox and every other browser, and force a restart through your device management so the update actually applies.
- Ask your hosting and IT providers two questions. How quickly do they apply kernel updates, reboot included? And who on their side would see a low-privilege account turn into root?
- Rescan your own code with each new model. OpenAI's research shows new models find what older ones missed in the same code. If you write software, have it looked at again, by a person or a tool, when the tools get better.
If you'd like a second opinion on your patching, isolation or exposure, our cyber security consulting and penetration testing teams can help, and our managed Linux team patches servers for a living. Or just get in touch.
Frequently asked questions
Can attackers use GPT-6 Cyber?
Not legitimately. Access is limited to vetted customers, API access is locked to approved networks by default, and OpenAI says it monitors for misuse. It also says it sees growing attempts by malicious groups to misuse or copy its models. The bigger worry is that other models, including ones anyone can download, catch up.
Is my Android phone affected?
Android runs on the Linux kernel, so some of these fixes will reach phones over time, through Google's monthly security updates and your phone maker. Install them when they arrive. Phones that no longer get updates are the ones to worry about.
Is this good news or bad news?
Both. The flaws were already there, and now hundreds of them are being fixed. But it shows how cheap finding them has become, and attackers will have tools like this too. OpenAI's announcement says it plainly: there is "a limited window" to use these models to secure software first.
Related reading: GLM-5.3: Anyone Can Now Download an AI That Breaks Into Software, OpenAI Paused Frontier Training, Then Red-Teamed Its Own Containment and AI Agents Built a Message Board and Broke Out.
Photo credits. All images used under the Unsplash License — free for commercial use, attribution appreciated:
- King penguins on the march — Yomex Owo
- Red fox — Cody Boileau
Sources
- GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents — OpenAI, 5 November 2026. Benchmarks, the Critical rating, the Firefox findings, access and safeguards.
- A thousand fewer paths to root — OpenAI, 4 November 2026. The Linux kernel pipeline, its costs, the July comparison run and the maintainer response.
- GPT-6 Cyber system card — OpenAI. Full evaluation results.
- Mozilla Foundation Security Advisory 2026-118 — Mozilla, 27 October 2026. Security fixes in Firefox 159.
- OpenAI's new GPT-6 Cyber AI finds 81 Firefox flaws, 948 Linux bugs — BleepingComputer, 6 November 2026.
Was this useful?
Thanks — noted.
Share
Petr Červenka
Petr is the founder and lead developer at Nano Solutions, a Perth-based custom software firm. With over a decade of experience building enterprise platforms for government and private sector clients, he leads delivery of complex projects across Australia.
Connect on LinkedIn