OpenAI has unveiled a research prototype that uses an upcoming cybersecurity model to find and verify vulnerabilities in the Linux kernel, create exploits, and propose patches. According to the company, during a test run in October, it generated exploits for 948 distinct flaws through which a local user without special privileges could gain root access. Human researchers then review the results.
From Code Selection to Verification
The system works in several consecutive steps. According to OpenAI, scout agents initially selected 5,034 areas for review, covering around 2.8 million lines of code and thus approximately seven percent of the kernel. Additional agents analyzed these areas within twelve hours and, after removing duplicates, produced 2,308 potential findings.
During the subsequent verification process, the company focused on findings that were accessible to low-privileged users, required no special permissions, and exhibited measurable memory corruption. For each case, an agent created and refined a proof of exploitation in an isolated virtual machine using the KASAN tool.
Exploits and Proposed Patches
According to OpenAI, the prototype dynamically reproduced 1,724 vulnerabilities in the controlled research environment. To assess their exploitability, it used isolated virtual machines with the protections of a Linux distribution configured with default settings. This was intended to ensure that the tests did not depend on disabled security features.
The agents also created 291 proofs of information-leak exploitation and 1,169 functional local privilege-escalation exploits. After analyzing the variants, OpenAI prepared 948 distinct patches and used this figure as the number of unique vulnerabilities of this type. The October run cost a total of $1,213,430 in API credits.
According to OpenAI, the output for human researchers includes a root-cause analysis, a minimal proof that triggers a crash, a proposed patch, and a complete exploit. These materials are intended to help them assess the severity of a flaw and forward fixes to the project maintainers.
Comparison with Older Model and kernelCTF Test
OpenAI compared the October run with a scan performed by the GPT-5.6 Sol model in July. According to the company, the older model found and exploited 166 distinct local privilege-escalation vulnerabilities; the upcoming cybersecurity model found and exploited 948 in the later run. Both scans examined the kernel's then-current main development branch.
In a separate test in a customized environment of the kernelCTF program, the system achieved 25 proofs of information leaks and 21 local privilege-escalation exploits, according to OpenAI. The company removed the SLAB_VIRTUAL protection mechanism from the configuration because it was incompatible with the kernel version used.
OpenAI also analyzed a sample of 166 agent-generated exploits. To the best of the company's knowledge, ten new exploitation techniques were used in them. The company discussed these with external Linux kernel exploit experts, who also said the techniques were new to them.
Human Oversight and Patch Disclosure
Human review uncovered a problem during an early run. According to OpenAI, of 200 seemingly successful exploits, 20 had obtained root privileges through a different, previously unknown vulnerability than the one assigned to the agents. The company subsequently tightened the task specification and added a check to determine whether the exploit used the specified vulnerability.
During the ten days taken by the scan and manual analysis, 61 of the privilege-escalation vulnerabilities found had already been fixed in the main branch, according to OpenAI. Patches were ready for another 43, while 299 had been publicly reported or discussed.
Following the recommendation of the Linux kernel community, OpenAI has already sent hundreds of reports and proposed patches for critical vulnerabilities. The company is working with members of the Patch the Planet initiative to gradually disclose the remaining fixes.
