An automated system built by OpenAI found 1,724 real vulnerabilities in the Linux kernel and produced 948 working exploits that let an unprivileged local user gain root access. The pipeline, powered by an unreleased OpenAI cyber model, scanned millions of lines of code in mid-October 2026, at a total API cost of $1,213,430.

Psst!Add AlexTech.ai asPreferred Source on Google

The Linux kernel runs on billions of Android phones, cloud servers and desktop PCs, and with over 40 million lines of code it exposes a huge attack surface through system calls and drivers. OpenAI says vulnerability discovery is becoming easier to scale, but prioritization and validation remain operational bottlenecks.

official image of OpenAI Cyber Model

Image: OpenAI — https://openai.com

How the pipeline works

The process has five stages. Scout agents selected 5,034 audit targets covering about 2.8 million lines of code, roughly 7% of the kernel, at a cost of $3,426 in credits. Analyst agents produced 2,308 candidate findings in 12 hours, for $316,549. Because static analysis produces many false alarms, Triage agents wrote crashing test programs and ran them in isolated virtual machines with the Kernel Address Sanitizer active. Within 30 hours they reproduced 1,724 vulnerabilities, 74.7% of the candidates, for $73,330. Only 3.1% of unconfirmed cases came from wrong model logic; the rest needed race conditions that were hard to trigger.

The next stage narrowed the confirmed cases to 313 possible information leaks and 1,238 root escalation candidates. ExploitDev agents tested them on standard desktop virtual machines with modern defenses such as KASLR, SMEP, SMAP and HARDENED_USERCOPY enabled. They produced 291 information-leak proofs of concept and 1,169 working root exploits, which after variant checks map to 948 distinct vulnerabilities.

Human review

In an early test, reviewers noticed that 20 of the 200 exploits reported as working gained root by triggering a different bug than the one assigned. The team added strict path checks to confirm each exploit used its exact target. The pipeline then generates patches, which human researchers review before sending them to kernel maintainers.

Costs and comparison with bounties

The average cost per unique root exploit was $1,279, while building a working exploit from a confirmed bug cost about $359 on average. For comparison, hacking contests such as Pwn2Own and TyphoonPWN pay $20,000 to $70,000 for one kernel root exploit. At those rates, the 948 exploits would be worth between $18.9 million and $66.3 million, or 16 to 55 times the cost of the scan.

OpenAI also tested the model against an adapted Google kernelCTF target with a smaller attack surface. Over 112 hours and $32,936 in credits, the pipeline produced 25 information leaks and 21 root exploits. If accepted by the program, those flaws would earn roughly $1.5 million in bounties.

New techniques

In a sample of 166 exploits, researchers identified at least 10 Linux kernel exploitation techniques not previously documented. Outside kernel experts confirmed they were new to them too: two new elastic objects for cache spraying, two new page grooming tricks, five data-only root attacks and one type confusion method. An agent auditing system call code also exploited a crash in the QEMU TCG emulator.

What it means for defenders

During the ten-day scan window, outside contributors patched or staged fixes for over 100 kernel flaws, while nearly 300 more bugs were discussed in public forums. The authors warn that an unpatched local user account is now functionally equivalent to root access for a well-resourced attacker. They call for automated patch generation as a standard tool, urge fast kernel updates, and recommend isolating processes with sandboxes such as Firecracker or gVisor.

OpenAI's earlier work

The research follows earlier work. In June, OpenAI's GPT-5.5 Cyber had found kernel bugs under the Patch the Planet project, building 24 root exploits to verify its reports. In July, a scan with GPT-5.6 Sol found and exploited 166 unique flaws for $513,690 in API credits. The October run with the newer model found more than five times as many unique root bugs.