In a single run in mid-October, an OpenAI system found 948 working ways to turn an ordinary Linux user account into root. Not theoretical flaws — 948 finished, tested exploits, each one lifting an unprivileged user to full control of a stock desktop Linux system. OpenAI published the details on 5 November, and the numbers underneath the headline are the interesting part.
The work sits inside the company's Patch the Planet initiative. A five-stage pipeline, running on an upcoming cyber model now named ChatGPT Cyber, scanned the latest Linux kernel master branch and flagged 2,308 deduplicated candidate findings. Nothing reached maintainers unsupervised: human researchers reviewed every exploit, patch, and severity rating first, and hundreds of reports and patches have already gone upstream.
How the hunt worked
Scout agents worked through the kernel folder by folder, flagging code reachable by low-privileged users and compressing more than 40 million lines of code into 5,034 audit targets. Analyst agents, running a workflow modelled on OpenAI's Codex Security plugin, scanned that filtered set and returned the full candidate list within half a day.
Triage agents then wrote proof-of-concept exploits and fired them inside isolated virtual machines running a kernel instrumented with Kernel Address Sanitizer, iterating until a crash reproduced or the trigger paths ran dry. That stage confirmed 1,724 real vulnerabilities, which a screening agent split into information leaks and privilege escalations before exploit-development agents built working exploits and variant analysis consolidated them into unique vulnerabilities, each with its own candidate patch.
Only 74.7 percent of the candidates survived dynamic validation, and model error accounted for just 3.1 percent of the failures. The targets were not weakened builds: the stock desktop distribution ran with SMEP and SMAP enabled alongside KASLR and HARDENED_USERCOPY, the same threat model used at Pwn2Own and TyphoonPWN.
What it cost — and what the haul is worth
The full October campaign cost $1,213,430 in API credits, or $1,279 for every working root exploit — an upper bound, OpenAI says, since better harness engineering could cut spend at every stage.
Valued against historical contest bounties, the haul carries a hypothetical valuation of $66.3 million at Pwn2Own and TyphoonPWN rates for Linux kernel privilege escalations — roughly 16 to 55 times what the campaign cost. Google's hardened kernelCTF environment provided a second proving ground: the pipeline produced 21 privilege-escalation exploits there in one campaign, nearing the 26 confirmed zero-days accepted from human submitters across all of 2026.
The scale-up is just as striking. In June, GPT-5.5 Cyber built 24 exploits to validate the flaws it found. In late July, GPT-5.6 Sol produced 166. The October run returned more than five times the July count, even though the kernel tree had absorbed months of additional hardening in between. The agents also deployed at least 10 exploitation techniques that third-party kernel experts confirmed were new to them.
The audit was honest about its own misses
What makes this disclosure worth taking seriously is how hard OpenAI audited its own success claims. An early run logged 200 successful exploits by the system's count. Human reviewers went back through every one of them and found 20 had gained root through a different zero-day than the one assigned, so OpenAI tightened the task constraints and added an independent check tying each final exploit to its assigned bug.
That same care shows up in how quickly the findings moved. Within 10 days, 61 of the found privilege escalations had been patched upstream, staging patches covered dozens more, and public discussion was open on hundreds of the remainder. Proving which AI model answered you is getting its own tooling too — Anthropic's FlashProve is built around proving which AI model answered you — but the patching still depends on humans shipping the fixes.
ChatGPT Cyber and the Daybreak programme
TechCrunch reports the model is called ChatGPT Cyber, available to the highest tier of its Daybreak programme, Daybreak Red, which serves offensive cybersecurity researchers. The defender-facing tier, Daybreak Blue, gives cybersecurity professionals access to frontier models — GPT-5.6 Sol among them — and it is already producing results in ordinary hands.
Max Justicz, chief technology officer of Charge Robotics, used GPT-5.6 Sol through Daybreak Blue to find a severe vulnerability in code running on Ubuntu servers. He had previously found a bug in the same codebase, aimed the model at it, and had a working exploit within two hours. He reported it to the Debian Project, which issued a patch. There has been no known evidence of exploitation — but as Justicz put it, "we should expect some extremely crazy bugs to be found and exploited over the next few months."
OpenAI has also announced a $1 billion initiative to support cybersecurity professionals working on critical infrastructure like the electrical grid or water systems. The same model found 81 novel vulnerabilities in Firefox. The same agent machinery that powers OpenAI Dots, the company's always-on assistant, is now hunting vulnerabilities at machine speed — and the question is who gets to use it first.
Why the patching clock matters now
OpenAI's own conclusion is blunt: "For the near- and medium-term, having access to a low-privileged account on an operating system that isn't up-to-date on its latest available patches should be considered functionally equivalent to root access for a well-resourced threat actor."
The gap between finding a vulnerability and weaponising it is collapsing, and the cost of doing it at scale is now measured in hundreds of dollars per exploit, not months of expert labour. Patching promptly used to be good hygiene. It is now the only thing standing between an unpatched machine and root access.
Three things will tell us how this plays out: whether defenders get access to tools like ChatGPT Cyber fast enough; whether other labs publish hunts of their own; and whether the 948 patches land before the techniques behind them spread. OpenAI says there is a narrow window to patch the infrastructure the world depends on. The window is real — and it is getting narrower.




