AI detects 948 privilege escalation flaws in Linux

An OpenAI AI system identified and exploited 948 privilege escalation vulnerabilities in the Linux kernel, the core component that powers Android, servers, supercomputers and millions of devices. The analysis also generated candidate patches, which experts must review before they can be incorporated into the project.

The test took place in mid-October 2026 using a cybersecurity model that has not yet been announced. In total, the system analyzed 2,308 potential flaws, dynamically reproduced 1,724 and created proof-of-concept exploits for 291 information disclosure vulnerabilities and 948 flaws that could allow an unprivileged local user to gain administrator access, known as root in Linux.

The process cost $1.21 million in API credits, or about $1,279 for each identified privilege escalation. OpenAI says the main goal was to measure the model's capabilities, not optimize cost, so it considers these figures an upper bound that could be reduced through better system engineering.

From thousands of alerts to verified flaws

The system did not indiscriminately review the kernel's more than 40 million lines of code. It first selected 5,034 targets related to parts of the system that users with limited permissions can reach. Those targets represented about 2.8 million lines, approximately 7% of the total code.

Different AI agents then searched for potential errors and produced 2,308 findings in less than 12 hours. A static finding does not by itself prove that a real vulnerability exists, so another component attempted to reproduce each case in isolated virtual machines.

Within 30 hours, the system confirmed 1,724 cases, around 74.7% of the candidates. Most of the remaining cases required conditions that were difficult to recreate, such as races between processes or memory allocation failures. OpenAI says only 3.1% were the result of incorrect analysis by the model.

The next phase measured what each flaw could do. The agents tested the cases with the standard protections used by a Linux distribution instead of running a deliberately weakened system. They produced 1,169 working local privilege escalation exploits, but further analysis and comparison between variants reduced the figure to 948 distinct vulnerabilities.

AI also proposed the patches

For each validated vulnerability, the system generated a candidate patch, a root-cause analysis, a minimal reproduction of the flaw and a full demonstration of its impact. Human researchers reviewed the results and began submitting reports and patches to kernel maintainers, following the recommendations of the Linux community.

That does not mean all 948 problems have been fixed. OpenAI reported that during the ten days of analysis, 61 local privilege escalations had already been fixed, 43 had patches in preparation and another 299 were being publicly discussed by the community. Some findings may have overlapped with other researchers' work or become outdated as the code changed.

Human review remains essential. In an early version of the experiment, 20 of 200 purported exploits had gained administrator privileges by exploiting a different vulnerability from the one they were supposed to test. The team changed the process and added independent checks to verify that each result actually used the assigned flaw.

A leap from the previous test

In July 2026, another OpenAI model found and exploited 166 unique local privilege escalation vulnerabilities at a cost of $513,690. The October test reached 948 cases at a total cost of $1.21 million, despite also analyzing the kernel's most recent branch at the time.

The researchers also say that, while studying a sample of 166 exploits, they identified at least 10 exploitation techniques they did not know about. They did not publish the technical details to avoid facilitating attacks against real systems.

In a second evaluation, conducted in a hardened environment based on kernelCTF, the system obtained 21 local privilege escalation exploits and 25 information disclosure proofs of concept. That campaign cost about $32,936. A comparison with historical rewards from security programs suggests that the potential value of the flaws would far exceed the cost of finding them, although this does not mean all of them would have been accepted or paid for.

What changes for you

Linux is present in many places even when you do not see it directly: Android phones, cloud services, routers, enterprise systems and connected devices. When a vulnerability appears in the kernel, the potential reach can be enormous, but the risk depends on whether an affected version is in use, whether the attacker has local access and whether the system has already received the patch.

The practical takeaway is simple:

• Keep your operating system, kernel and applications up to date.
• Do not treat a low-privilege local account as sufficient protection on outdated machines.
• Use isolation and sandboxes whenever possible, especially on servers and in environments that run third-party code.
• Keep in mind that AI-generated patches need testing, security review and checks against variants of the same flaw.

OpenAI says large projects will likely need several rounds of analysis and remediation as models improve. The race does not end when a vulnerability is found: maintainers must confirm the problem, fix it, distribute the update and get users to install it. The next challenge will be making that cycle fast enough that finding flaws becomes less profitable for attackers than fixing them in time.