OpenAI Finds 948 Privilege Escalation Flaws in the Linux Kernel

OpenAI has published the results of a large-scale analysis of the Linux kernel carried out with AI agents. According to the company, the effort cost more than a million dollars and ended with 948 distinct privilege escalation vulnerabilities identified in the kernel. Patches have been prepared for these flaws; some have already been fixed, while the rest are being worked through.

The company says it designed the process as a five-stage pipeline: vulnerability discovery, dynamic validation, exploitability assessment, patch generation and human review. The total cost of the project is reported as roughly $1.21 million in API credits.

How the five-stage pipeline worked

In the first stage, AI agents identified the security-relevant parts of the kernel and scanned them for potential weaknesses. Within about 12 hours, more than 2,300 candidate findings emerged. Other agents then generated proof-of-concept (PoC) code for each candidate and tested it. In this validation stage, which took about 30 hours, roughly 75 percent of the findings, or 1,724, could be reproduced. Only 3 percent of the remaining findings are said to have resulted from faulty analysis by the model.

In the next step, the agents ranked the flaws by impact. This stage produced 1,169 working local privilege escalation exploits and 291 PoCs demonstrating information leaks. After duplicates and false positives were filtered out, 948 distinct privilege escalation vulnerabilities remained, and patches were prepared for them.

Patches and coordination with kernel developers

For patch generation, OpenAI says it gave the model the relevant source code along with best practices for writing patches. These included running static checks on the patch, following the coding conventions of nearby code, and performing variant analysis to make sure the fix was complete.

The company's researchers manually reviewed every exploit and patch before contacting Linux kernel developers. At the Linux community's direction, the flaws were prioritized. Some have already been patched, and the rest are reportedly being addressed.

Not all findings were new, but some techniques were

Not every finding was entirely new. By the time the analysis was complete, kernel developers had already patched 61 of the identified privilege escalation flaws, and patches for another 43 were in the works. Roughly a third, close to 300 of the flaws, had also been publicly reported and discussed.

The team also examined 166 exploitation techniques and found that 10 of them had not been documented before. These include manipulation of kernel memory, page grooming and privilege escalation through data-only attacks. OpenAI said it will not share details of these techniques until the flaws are closed, so as not to give attackers an advantage.

OpenAI explains its choice of Linux this way: the kernel underpins Android devices, cloud infrastructure and supercomputers, and its codebase of more than 40 million lines offers a broad attack surface through system calls, networking, file systems and device drivers. The company wanted to measure how well frontier models can find vulnerabilities and demonstrate their impact in a codebase of this scale.

The report refers to the model that found more than 900 flaws only as an upcoming cyber model, without naming it, and compares it with GPT 5.5-Cyber, which is said to have developed 24 local privilege escalation exploits. The study may therefore double as a showcase for the company's new cyber model, but the results are striking enough that they cannot be dismissed.