OpenAI tests an AI pipeline against Linux flaws
OpenAI says it detected 948 local privilege escalation vulnerabilities in Linux with a prototype pipeline. The experiment, conducted in October 2026, mainly measures a scale change for automated offensive research.
OpenAI announced on November 4, 2026, that it had found 948 local privilege escalation vulnerabilities in Linux using a prototype pipeline based on a future cyber model. The experiment, conducted in mid-October 2026 on isolated virtual machines, demonstrates that AI can now chain discovery, exploitation, and patch proposal on a massive codebase. The identified flaws could grant root access to a non-privileged user. And on Linux, that’s never the kind of detail you tuck away in a drawer.
A test designed to measure the full chain, not just detection
OpenAI did not test an artificially weakened system. The trials took place on a standard desktop operating system, with SMEP, SMAP, KASLR, and HARDENED_USERCOPY enabled, in isolated VMs. The choice of Linux is also explained by its size: the kernel exceeds 40 million lines of code and powers Android smartphones, cloud infrastructures, and supercomputers.
The pipeline links several steps: target selection, code analysis, dynamic validation, exploitability estimation, patch generation, then human review. Scout agents identify areas accessible to low-privilege users and considered security-sensitive. Analyst agents then examine these targets, followed by Triage agents generating and refining proofs of concept in VMs with Kernel Address Sanitizer, to keep execution traces rather than relying solely on static analysis.
From vulnerabilities to the work stack, then to the patch stack
Each complete research file contains a root cause analysis, a minimal proof of concept causing a crash, a candidate patch, and a full exploit. OpenAI says it selected 5,034 audit targets covering about 2.8 million lines, or 7% of the kernel. Analysts produced 2,308 candidate findings deduplicated in 12 hours, then dynamic validation reproduced 1,724 vulnerabilities in 30 hours.
The next phase was even more concrete: 291 proofs of concept for information leaks and 1,169 functional local exploits, according to OpenAI. The variant analysis step then produced 948 distinct patches, which the company counts as 948 unique vulnerabilities. These patches remain candidates subject to maintainer review, not merged upstream patches.
Why the figure of 948 doesn’t tell the whole story
OpenAI specifies that several results concerned the same vulnerabilities in different forms. Hence the gap between the number of raw exploits and the final total of unique vulnerabilities. The company also says it identified 61 LPEs already fixed upstream, 43 with patches in preparation, and 299 already public or discussed at the end of the ten-day scan and manual analysis. The message to maintainers is clear: results must be revalidated on current code, including subsequent merges, to avoid duplicates or already addressed vulnerabilities.
In an initial cycle, OpenAI also spotted a methodological issue more embarrassing than it seems: 20 “successful” exploits out of 200 obtained a root shell via another 0-day than the assigned vulnerability. The team then tightened guidelines and added independent verification of the exploitation path. Even a root shell can lie a little, apparently.
Patch quality matters as much as the exploit
Patches generated too quickly could be correct in substance while adding unnecessary code, degrading performance, deviating from project conventions, or imperfectly covering variants. OpenAI says it improved patch quality through additional instructions and iteration, but human researchers continue to validate the exploit, patch, and proposed severity before any submission to maintainers.
This point matters for security teams as well as Linux maintainers. The real bottleneck is no longer only in generating leads, but in proving that an attack path is real, reproducible, and fixable without unnecessary noise. This also distinguishes defensively exploitable research from a flood of reports difficult to sort.
How much does it cost, and what do comparisons say?
OpenAI estimates that the mid-October 2026 campaign cost 1.213 million dollars in API credits, or 1,279 dollars per unique LPE. By comparison, a GPT-5.6 Sol session conducted in late July 2026 found 166 unique LPEs for 513,690 dollars. The announced gain thus exceeds a factor of five on new vulnerabilities, despite three months of hardening between the two scans.
The company also indicates having observed, on a sample of 166 exploits, ten techniques considered new by its team and by third-party kernel exploitation experts. Technical details were deliberately withheld for security reasons. Another experiment, adapted to kernelCTF, produced 25 proofs of information leaks and 21 LPE exploits in 112 hours for about 32,936 dollars in API credits. The test had to remove the SLAB_VIRTUAL mitigation, deemed incompatible with the targeted kernel, and the amount of about 1.5 million dollars in reward remains conditional on acceptance of the 21 vulnerabilities. Google kernelCTF rules provide 71,337 dollars per eligible exploit, with a stability rate of at least 70% over 20 runs.
Key points
- 948 unique LPEs detected in Linux, according to OpenAI, November 4, 2026.
- 5,034 audit targets, about 7% of the kernel.
- 1,724 vulnerabilities reproduced in 30 hours.
- Trail of Bits and OpenAI launched Daybreak on June 22, 2026.
- 166 unique LPEs found in July 2026 for 513,690 dollars.
- Google kernelCTF: 71,337 dollars per eligible exploit.
By the numbers
- 948 — unique local privilege escalation vulnerabilities, Linux, mid-October 2026, OpenAI.
- 2.8 million lines — audited scope, about 7% of the kernel, OpenAI.
- 1.213 million dollars — announced cost in API credits, October 2026 campaign.
- 166 unique LPEs — GPT-5.6 Sol session, late July 2026, OpenAI.
- 21 exploits — result of kernelCTF adaptation, 112 hours, OpenAI.