OpenAI has uncovered 948 vulnerabilities in the Linux kernel, which can allow limited users to gain full system control.
OpenAI has announced that its AI agents have uncovered 948 unique vulnerabilities within the Linux kernel, allowing users with limited permissions to gain complete control over systems.
Our security researchers built an autonomous cyber vulnerability discovery, exploitation, and patching pipeline. In a single run with an upcoming cyber model, it validated and generated patches for 948 local privilege escalation vulnerabilities in the Linux kernel at a total API cost of $1.2 million.
— OpenAI (@OpenAI) November 4, 2026
The AI prototype autonomously analyzed source code, identifying weaknesses, testing their exploitability, and generating fixes. This was achieved using a specialized cybersecurity model, details of which have not been fully disclosed by OpenAI.
Earlier in October, the AI detected 2,308 potential vulnerabilities and successfully reproduced 1,724 of them in isolated virtual environments. Out of these, data leakage was demonstrated in 291 cases, culminating in 1,169 exploit scenarios to gain administrator rights. After consolidating scenarios based on common errors, researchers concluded there were 948 unique vulnerabilities, each paired with an individual patch.
The testing was conducted on a standard desktop version of Linux, which had basic kernel protection. Importantly, these vulnerabilities require local device access to exploit.
The research cost about $1.21 million, averaging approximately $1,279 for each identified flaw. Compared to July's tests with the GPT-5.6 model, which discovered only 166 unique vulnerabilities for $514,000, the new system proved significantly more effective despite recent improvements to Linux's defenses.
Novel Hacking Techniques Discovered
OpenAI researchers analyzed a sample of 166 scenarios and uncovered at least ten new hacking techniques previously unknown to experts. These findings were also shared with third-party Linux security specialists, who reported encountering similar unfamiliar methods.
Some of the new techniques included innovative ways to cache kernel objects, enhance attack reliability, manipulate data types, and gain administrator rights by directly modifying memory contents. Due to safety concerns, OpenAI has not disclosed specific technical details.
The AI system was also tested in Google’s kernelCTF environment, which utilizes enhanced security measures. Here, it generated 21 working attack scenarios and demonstrated data leakage in 25 instances, with this specific testing costing around $33,000.
Patching Vulnerabilities
While patches have been created, each needs validation and approval from the maintainers of the Linux kernel. Following community guidance, OpenAI prioritized critical vulnerabilities and has already submitted hundreds of reports, with plans to share remaining findings through the Patch the Planet initiative.
By the end of the ten-day investigation, 61 vulnerabilities were closed in the main Linux repository, while 43 patches were in intermediate branches and 299 were under public discussion, with assistance from external security researchers.
OpenAI emphasized that AI not only speeds up the identification of vulnerabilities but also the development of attacks. Researchers believe that as these systems evolve, there will be a growing need to automate the release of patches and enhance the security of critical software.
This follows OpenAI's previous autumn discovery of two critical vulnerabilities in PHP, which led to collaboration among developers, security researchers, and infrastructure companies to address the issues.
