OpenAI AI Agents Discover 948 Local Privilege Escalation Vulnerabilities in Linux Kernel
November 7, 2026 · 2 min read
Also published in Português (Brasil)
OpenAI announced on November 4, 2026, that AI agents using a cybersecurity AI model under development autonomously discovered and verified vulnerabilities in the Linux kernel.
During tests conducted in mid October on the latest master branch of the Linux kernel, the agents dynamically reproduced 1,724 vulnerabilities. This process resulted in the generation of 291 proof of concept codes for information leakage and 948 exploits for local privilege escalation, which allow general users to obtain root administrator privileges. The agents utilized an automated pipeline to handle discovery, verification, exploit creation, and the generation of patches.
To begin the process, the agents identified 5,034 target areas totaling approximately 2.8 million lines of code, which is about 7 percent of the total kernel source code of 40 million lines. After approximately 12 hours of static analysis based on a threat model, the agents detected 2,308 vulnerability candidates. These were then verified using programs executed on isolated virtual machines with the Kernel Address Sanitizer enabled.
Analysis of 166 sampled exploits by OpenAI researchers revealed the use of at least 10 novel Linux attack methods previously unknown. The total cost for API credits during this investigation was 1,213,430 dollars, averaging about 1,279 dollars per privilege escalation vulnerability. OpenAI noted that similar findings in security contests like Pwn2Own or TyphoonPWN typically command rewards between 20,000 and 70,000 dollars per case, meaning the total value for 948 cases would be between 18.9 million and 66.3 million dollars.
Earlier in late July 2026, GPT-5.6 Sol discovered 166 unique privilege escalation vulnerabilities with a total API credit cost of 513,690 dollars, or about 3,095 dollars per case. The company observed that as models improve, they require less complex orchestration and less hand holding. While GPT-5.5 Cyber required external frameworks for skill reinforcement to apply exploit techniques, GPT-5.6 Cyber and newer models achieved high success rates with simpler configurations.
During the 10 day analysis period, the Linux community integrated 61 local privilege escalation fixes into the kernel, with 43 staging patches prepared and about 300 others under public discussion. OpenAI researchers have sent hundreds of reports and proposed fixes to kernel developers, prioritizing severe issues. The company is also working with organizations in Patch the Planet to submit remaining reports and patches.