November 5, 2026
OpenAI says a pipeline built on an unreleased cyber model found, proved, and drafted patches for 948 distinct Linux kernel vulnerabilities that let an ordinary local user take root access — all in a single research run in mid-October. Why it matters: automated exploit-building at this scale changes the math for every team that defends Linux, which runs most of the cloud, Android, and the world’s supercomputers.
What happened
In a research post published November 4, titled “A thousand fewer paths to root,” OpenAI described a five-stage pipeline — discovery, dynamic validation, exploitability assessment, patch generation, and human review — run against the latest Linux kernel. Scout agents picked 5,034 audit targets covering about 2.8 million lines of code, roughly 7% of a kernel that runs to more than 40 million lines. Analyst agents produced 2,308 candidate findings within 12 hours, and triage agents working in isolated virtual machines reproduced 1,724 of them as real vulnerabilities within 30 hours. OpenAI says only 3.1% of the candidates turned out to be incorrect model analysis.
The details
The exploit stage produced 1,169 working local privilege-escalation exploits plus 291 information-leak proofs of concept. After variant analysis merged exploits that shared a root cause, that left 948 distinct vulnerabilities, each with a candidate patch written by the pipeline. The test target was not a weakened setup: the virtual machines ran a stock desktop operating system with standard protections — SMEP, SMAP, KASLR, and HARDENED_USERCOPY — switched on. The full run cost $1,213,430 in API credits, about $1,279 per unique exploit, a figure OpenAI itself calls an upper bound rather than an optimized price.
The jump from earlier runs is the story. In June, under OpenAI’s Patch the Planet program, GPT-5.5 Cyber produced 24 such exploits. A late-July run with GPT-5.6 Sol found 166, at a cost of $513,690. The October run found more than five times that — despite three months of kernel hardening in between. In a sample of 166 exploits, OpenAI says its agents used ten exploitation techniques that its researchers and outside kernel experts consulted believed were new. Full technical details were withheld for safety reasons, and all of these numbers are OpenAI’s own; they have not been independently verified.
Where the fixes stand
By the end of the ten-day scan and manual analysis, 61 of the flaws had already been patched upstream, 43 had patches in staging, and 299 were being reported and discussed publicly. OpenAI says it has sent hundreds of reports, starting with the most critical, and will send the rest gradually with its Patch the Planet partners so volunteer maintainers are not buried. Human researchers review each exploit, patch, and severity rating before anything goes to a maintainer. The 948 figure counts validated exploits with candidate patches — not fixes that have already shipped.
Why it matters
For anyone running Linux servers, the practical takeaway is OpenAI’s own blunt guidance, as reported by security outlets: treat an unpatched machine with a low-privilege account on it as effectively root-accessible to a well-resourced attacker. The bottleneck in security research is shifting from finding bugs to proving, patching, and reviewing them — and that review burden lands on maintainers. This is “everything AI, tested” in the most literal sense: the capability is real enough to measure in dollars per working exploit, and defenders now have to plan around it.
Frequently asked questions
Did OpenAI release the exploit code?
No. OpenAI withheld the technical details for safety reasons and is disclosing the flaws to kernel maintainers through its Patch the Planet program, starting with the most critical.
What is a local privilege-escalation vulnerability?
It is a flaw that lets someone who already has an ordinary, low-privilege account on a machine upgrade themselves to root — full administrative control. An attacker still needs that initial foothold first.
Are the 948 patches already in Linux?
No. Each exploit came with a candidate patch drafted by the pipeline, but maintainers decide what ships. At the end of OpenAI’s analysis, 61 flaws were patched upstream and 43 more had patches in staging.
How much did the run cost?
OpenAI reports $1,213,430 in API credits — about $1,279 per unique exploit — and describes that as an upper bound, since the run was designed to measure capability, not to minimize cost.
Sources: OpenAI (“A thousand fewer paths to root,” November 4, 2026), Unite.AI, GBHackers, The Hacker News