OpenAI Says Its AI Found and Exploited 948 Linux Kernel Flaws in a Single Run

OpenAI on Wednesday detailed a pipeline that employed an upcoming cyber-focused artificial intelligence model to validate and patch 948 local privilege escalation vulnerabilities in Linux for $1.2 million in API costs. The findings were published in a research post titled “A thousand fewer paths to root.”

“Our prototype identified and analyzed 2,308 candidate Linux kernel findings, dynamically reproduced 1,724 vulnerabilities, and produced 948 local privilege escalation (LPE) exploits that could allow an unprivileged local user to gain root access,” OpenAI said. The run was conducted in mid-October 2026 against the latest Linux kernel master branch, in about 30 hours.

The flaws let someone who already has limited access to a Linux machine, such as a basic user account on it, take full control of it. OpenAI said they worked on a standard Linux desktop with normal security protections turned on, including SMEP, SMAP, KASLR and HARDENED_USERCOPY, in isolated virtual machines.

OpenAI said its system works in steps. It finds risky code and looks for flaws, then tests them in a safe sandbox. It then checks whether they can be used in an attack and, if so, drafts a fix. The pipeline identified 2,308 candidate findings and reproduced 1,724 vulnerabilities. Its agents generated 1,169 working local privilege-escalation exploits, and variant analysis produced the final count of 948 unique cases, each with a candidate patch.

The total run cost $1,213,430 in API credits, or $1,279 per LPE exploit. OpenAI said the 948 flaws are hypothetically worth $18.9 million to $66.3 million, against the $1.2 million it cost to run the AI, going by what typical cash prizes are paid at hacking contests such as Pwn2Own. A July run with GPT-5.6 Sol had found 166 unique escalations for about $514,000.

OpenAI said its goal is defense, meaning finding these flaws before hackers do, so they can be fixed. It said it has sent hundreds of reports and patches for critical vulnerabilities to kernel maintainers and plans to distribute the remainder gradually with Patch the Planet partners. The candidate fixes still need review and integration before they become updates users can install, and the model behind the run has not been released.