The OpenAI company disclosed information about the identification of 1724 vulnerabilities in the Linux kernel, for 948 of which it was possible to create working exploits that allow a local user to gain root access in the system, and for 291 vulnerabilities it was possible to prepare exploits leading to information leakage. The cost of analyzing code and generating exploits using the GPT-5.5 and GPT-5.6 Cyber AI models is estimated at $1.21 million, which is $1,279 for each exploit. To compare, the reward for developing one similar exploit at the Pwn2Own and TyphoonPWN competitions ranges from $20,000 to $70,000 (profitability from 16 to 55 times).
The performance of the exploits was tested in standard Linux distributions with the SMEP (Supervisor Mode Execution Prevention), SMAP (Supervisor Mode Access Prevention) protection mechanisms enabled, KASLR (Kernel Address Space Layout Randomization) and HARDENED_USERCOPY.
When tested in a secure environment with a limited attack surface, as used in Google’s kernelCTF competition, 21 privilege escalation exploits and 25 information leakage exploits were successfully deployed (compared to 26 privilege escalation exploits successfully demonstrated in the kernelCTF competition for all of 2026).
A multi-stage process was developed to find vulnerabilities as part of the Patch the Planet initiative process. At the first stage, AI agents, through static analysis, selected the most suitable code sections for searching for vulnerabilities, which made it possible to narrow the search area to 2.8 million lines (7% of the code base). After this, with the help of AI agents, the selected code was analyzed and candidates for vulnerabilities were identified.
At the second stage, false candidates were eliminated by writing exploit prototypes and testing them in isolated virtual machines. After this, the exploit prototypes were brought to a full working state, capable of bypassing standard protection mechanisms, such as address randomization. At the next stage, AI agents generated patches to fix vulnerabilities, and at the final stage, each exploit and patch was verified by a human.
During the study, AI agents independently developed and applied at least 10 previously unknown techniques for exploiting the kernel and bypassing protections. In one case, the AI model identified and exploited a vulnerability in the QEMU emulator, which was used to isolate the test environment. The vulnerability in QEMU was found by an AI agent while attempting to exploit a vulnerability in the kernel code used to transfer control to system calls. The attempts ended with an emergency stop and the AI agent determined that the cause was a bug in QEMU and included this vulnerability in the report. At the next stage, the AI agent for developing working exploits prepared the corresponding exploit for QEMU.