OpenAI Uses Security AI Model to Audit the Linux Kernel, Generating 948 Local Privilege Escalation Exploits
On November 4, 2026, OpenAI announced that it used an in-development cybersecurity AI model to conduct autonomous vulnerability discovery on the Linux kernel, screening 2,308 candidate vulnerabilities from approximately 2.8 million lines of code, dynamically reproducing 1,724 of them, and generating 948 local privilege escalation (LPE) exploit codes and 291 information disclosure PoCs, while also producing fix patches. Researchers sampled and analyzed 166 exploit codes and found at least 10 previously unknown attack techniques. The total API cost of this investigation was $1,213,430, equivalent to about $1,279 per privilege escalation vulnerability; OpenAI said that converted at the per-item bounty of $20,000 to $70,000 in competitions such as Pwn2Own, the cost is only 1/16 to 1/55. OpenAI also mentioned that as model capabilities improve, complex orchestration and detailed instructions are no longer necessary, and recommended rescanning previously audited code with stronger models.
Why read this
SelectedAI score 85
The original text provides the full-process data for vulnerability discovery, exploit code, and patch generation, as well as a comparison of API costs per individual vulnerability, which can be used to assess the feasibility of automated security auditing.
Topics
Body · AI translation
On November 4, 2026, OpenAI announced the results of an effort in which an agent autonomously discovered and verified vulnerabilities in the Linux kernel using an AI model for cybersecurity that is under development.
A thousand fewer paths to root | OpenAI
The test was conducted in mid-October. In the Linux kernel (the latest master branch at that time), 1,724 vulnerabilities were dynamically reproduced, and 291 information-disclosure proof-of-concept (PoC) codes and 948 local privilege escalation (LPE) exploits that allow a general user to obtain administrator (root) privileges were generated. Through a pipeline that consistently automates everything from discovery and verification to exploit creation and patch generation, the agent not only demonstrated the vulnerabilities but also created proposed fixes.
Narrowing down privilege escalation vulnerabilities from 2,308 candidates
From the kernel's approximately 40 million lines of source code, the AI agent identified 5,034 target areas reachable from a general user and subject to security auditing, comprising about 2.8 million lines (about 7% of the entire kernel). As a result of performing static analysis on these over about 12 hours while understanding the target threat model, it detected 2,308 vulnerability candidates.
Next, for dynamic verification, the agent generated and executed a verification program for each candidate on isolated virtual machines. In a kernel environment with KASAN (Kernel Address Sanitizer) enabled on the virtual machines, the agent generated and executed verification programs, dynamically reproducing 1,724 vulnerabilities.
Among the reproduced issues, those that could lead to information disclosure or privilege escalation were selected and verified with standard desktop Linux protection features enabled; the agent then generated code reproducing 948 LPE exploits and 291 information disclosures. When OpenAI researchers sampled and analyzed 166 of the exploits created by the agent, at least 10 types of previously unknown (novel) Linux attack techniques were used.
Vulnerability discovery cost and economics
The total cost of API credits used in this investigation was $1,213,430, or about $1,279 per privilege escalation vulnerability. According to the company, when converted using the bounties offered in the past at security contests such as Pwn2Own and TyphoonPWN ($20,000 to $70,000 per item), 948 items would total $18.9 million to $66.3 million, reaching 16 to 55 times the API cost of this effort in a simple comparison.
In addition, during development of this pipeline in late July 2026, when GPT-5.6 Sol discovered 166 unique privilege escalation vulnerabilities, the total API credits were $513,690, or about $3,095 per item.
During OpenAI's investigation this time, even during the 10-day investigation and analysis period, fixes for 61 LPE-related issues were merged into the Linux kernel mainline by the community, and staging patches were prepared for another 43 (with about 300 more under public discussion). While expressing respect for the prompt response of these Linux maintainers, the company has established a system of reporting only after re-verification against the latest code, in order to avoid burdening developers with duplicate reports.
When reporting the Linux vulnerabilities this time as well, the company's researchers prioritized serious issues in accordance with the advice of kernel developers and sent hundreds of reports and proposed fixes before the announcement. For the remaining issues, too, they are cooperating with organizations participating in "Patch the Planet," which supports open-source vulnerability fixes, to proceed with reporting to developers and submitting fix patches.
In addition, as a finding from this attempt, it was noted that the smarter the model becomes, the less need there is for complex orchestration or excessive hand-holding instructions. In the experiments, GPT-5.5 Cyber required complex orchestration and skill augmentation (strong assistance from external frameworks) to apply exploit techniques, whereas GPT-5.6 Cyber (or the new model under development) was able to achieve a high success rate even with a simple configuration that omitted such orchestration.
Including these results and the earlier cost evaluation, the company recommends rescanning previously audited code as model performance improves. It states that using higher-performance models may make it possible to discover potential bugs that past models overlooked, and that it will henceforth advance an approach in which defenders comprehensively proceed with fixes ahead of attackers.