OpenAI says an unreleased model produced patches for 948 Linux privilege-escalation bugs
A post dated Wednesday describes one mid-October run, on a stock desktop kernel, that reproduced 1,724 bugs and wrote candidate patches for 948 ways an ordinary local account could reach root. The API bill was $1,213,430. In an earlier pass, 20 of 200 reported successes had used a different bug than the one assigned.
Deep dive6 minIn-depth
Priya Shah6 min read
- OpenAI
- Linux
- Security
- Agents

Photo: OpenAISource page
OpenAI said on Wednesday that a research pipeline, built around an upcoming cyber model it did not name, validated Linux kernel bugs and wrote candidate patches for 948 local privilege-escalation vulnerabilities. The post is dated 4 November. A local privilege escalation, as the post uses the term, is a flaw that could let someone who already has an ordinary account on a machine gain root, the account that can change the system. The mid-October run cost $1,213,430 in API credits. OpenAI divides that by the 948 and gets $1,279 per exploit. Human researchers, the post says, reviewed the output before reports went to maintainers.
The same post sets the run against a smaller one the company had already described. In June, under a program it calls Patch the Planet, GPT-5.5 Cyber found Linux kernel bugs and produced 24 local privilege-escalation exploits, used to validate and report them. Wednesday’s account says that work is now a pipeline: find candidate bugs, check that they can be made to happen, test whether they can leak information or reach root, draft a patch, and have a person read the result. The model in the October run is described only as upcoming.
What the October run counted
The prototype identified and analyzed 2,308 candidate findings and dynamically reproduced 1,724 of them. It then produced proofs of concept for 291 information leaks and 1,169 working local privilege-escalation exploits. Variant analysis, OpenAI says, reduced those exploits to 948 distinct vulnerabilities, and the pipeline wrote 948 patches. The virtual machines ran a stock desktop operating system, not a system with its protections turned off. OpenAI says that matches the threat model used at contests such as Pwn2Own and TyphoonPWN.
Those contests, the post says, have historically paid $20,000 and $70,000 for a Linux kernel local privilege-escalation exploit. Applied to all 948, the arithmetic is $18.9 million to $66.3 million, which OpenAI calls about 16 to 55 times the API cost of the run. The post labels that a hypothetical valuation. It is not money paid, and it uses bounty figures from earlier contests, not a price anyone offered for this set of bugs.
Where the credits went
A first set of agents, which OpenAI calls scouts, marked the code a low-privileged user can reach: 5,034 audit targets, about 2.8 million lines, roughly 7 percent of the kernel, at a cost of $3,426. Analyst agents then produced the 2,308 candidates in 12 hours, for $316,549. A validation stage tried to make each candidate happen inside an isolated virtual machine and reproduced 1,724 bugs in 30 hours, about 74.7 percent of the candidates, for $73,330. OpenAI says only 3.1 percent of the candidates were wrong because the model had assumed a condition that could not occur. Most of the rest needed higher privileges, or did not reproduce in the time allowed.
Testing what those bugs could do was the expensive stage. OpenAI puts it at $775,929 in API credits, after a screening pass at $44,194 that left 313 information-leak candidates and 1,238 privilege-escalation candidates. OpenAI says a successful exploit cost $359 in credits on average. A later line in the same section puts the 1,169 successes at $360 each and 69 unsuccessful attempts at $3,922 each. OpenAI says a few hard cases, including race conditions, pulled the failed-attempt figure up, and that stopping after an hour with no progress would have cut the bill. The post says the costs are closer to an upper bound than to a tuned one, because the run was built to measure the model, not to economize.
An early version of that stage reported 200 successful privilege escalations. After people looked, 20 of them had reached root through a different vulnerability than the one the agent had been assigned. OpenAI says it then tightened the task and added a separate check that the final exploit used the assigned bug. The miss is in the post for a reason the company states directly: a root shell, by itself, did not prove the assigned bug was the one that produced it.
What maintainers had already patched
The scan used the latest Linux kernel in mid-October. When the automated work and the manual analysis finished, 10 days on, 61 of the privilege-escalation bugs had already been patched upstream, 43 had staging patches, and 299 were being discussed in public. OpenAI says it has sent hundreds of reports and patches for the critical bugs, starting there on advice from the kernel community, and that it is working with Patch the Planet partners on the rest. It ties the effort to a commitment it calls Daybreak. The patches in the post are candidates for maintainers. The post does not say they are already in a kernel a user can install.
A second run used Google’s kernelCTF, a hardened kernel with a smaller attack surface. OpenAI says it dropped one mitigation, SLAB_VIRTUAL, because the kernel version it scanned was incompatible with that setting. The run produced 25 information-leak proofs of concept and 21 local privilege-escalation exploits, took 112 hours, and cost about $32,936, with discovery accounting for 83.7 percent of the API spend. The post compares the 21 with 26 confirmed zero-day privilege-escalation exploits that kernelCTF had accepted in 2026 at the time of writing. If those 21 were accepted, OpenAI says, the bounty would be about $1.5 million, roughly 45 times the scanning cost. The same run also hit a bug in QEMU’s tiny code generator. OpenAI says that flaw was already fixed in upstream QEMU and was still present in the build used for the experiment. It does not describe the bug.
On the same kind of harness, the post compares models. In late July, GPT-5.6 Sol found and exploited 166 unique local privilege-escalation bugs, at an end-to-end cost of $513,690. The October model’s 948 is the count OpenAI calls more than five times as many, each time on the kernel’s master branch as it stood, with three months of hardening between the scans. In a sample of 166 of the October exploits, OpenAI says the agents used at least 10 techniques it regards as new, and that outside kernel specialists shown those techniques had not seen them either. The post says it is withholding how they work.
A separate measurement in the post is about the harness, not the bug count. OpenAI says GPT-5.5 Cyber did not reliably pick known techniques on its own. Giving it a library of 103 of them, across 63 candidates, lifted a success rate from 76.2 percent to 88.9 percent. GPT-5.6 Cyber, with that library removed, beat the more heavily scaffolded older model. Putting the library back added 1.5 percentage points. OpenAI’s reading is that the stronger model needed less of that scaffolding, and that a harness should be measured again when the model changes.
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.
— OpenAI, “A thousand fewer paths to root,” 4 November 2026
That sentence is the company’s judgment, printed after the tables, not a figure from the October count. The model that produced the 948 does not have a public name. OpenAI says the point of publishing the costs is to show defenders what the work took, and it says it expects later harness work to bring those costs down.


