OpenAI’s New AI Pipeline Uncovers 1,724 Linux Kernel Vulnerabilities, Offering Both Promise and Peril

OpenAI has a new, unnamed AI pipeline in development that it has deployed to pick out some 1,724 Linux kernel vulnerabilities in well under two days. Of these, it also developed 948 working local privilege escalation (LPE) exploits.

While those have become expected numbers in the era of Mythos and similarly capable frontier models, the other part of the news here is that the AI pipeline also develops its own candidate patches for the exploits it digs up. And it did all of this at a token cost of a little over $1 million, a fraction of what digging up kernel vulnerabilities would have cost via standard research methods.

Linux kernel vulnerabilities being addressed in private

Mass discovery of vulnerabilities via AI in a matter of hours is nothing new, at least not since private Mythos testing began around April of this year. But while most testing partners are turning the advanced AI on their own internal systems to probe for weaknesses, OpenAI shelled out a little over $1.2 million in compute to probe the Linux kernel that underpins a great deal of internet services and operating systems.

While the numbers are concerning, the kernel vulnerabilities are not being publicly disclosed. OpenAI is working with the “Patch the Planet” initiative it founded and other partners to address these issues on a coordinated disclosure schedule, and indicates over 100 of the vulnerabilities have already been addressed. The AI pipeline provides a potential patching solution for each, but a human must review each one to ensure they are viable and do not break other components.

One key takeaway is thus that there remains something of a bottleneck to this process that is a disadvantage to the “good guy” defenders. The AI pipeline is sent to scan security-relevant portions of the Linux kernel, find and reproduce kernel vulnerabilities, create proof-of-concept exploits for them, and then create candidate patches for those exploits. In this case, it did all this in about 30 hours. However, there is still a human element at the end that must manually review the results and assign severity ratings.

AI pipeline could produce millions of dollars in savings

Once commercially available, the AI pipeline still seems to be something that would be highly beneficial to organizations in terms of efficiency and cost savings. OpenAI clearly sees this exercise as a $1 million investment in PR and indirect advertising, but the numbers they put forward cannot be ignored. The process breaks down to about $1,279 USD for not only the discovery of each of the kernel vulnerabilities, but also a potentially working patch delivered along with the news. In terms of cost alone, that is a massive savings over what usually runs at least tens of thousands of dollars to pay out a bug bounty for each vulnerability discovered by a researcher.

Of course, the pace of work must also be considered. The AI pipeline was also tested against the higher-security adapted kernelCTF⁠ environment, which offers substantially fewer avenues of attack. In these tests, the AI basically did in hours what security researchers have collectively spent the last 10 months doing: it uncovered 25 information leak PoCs and 21 LPE exploits, as compared to the 26 confirmed zero-days thus far uncovered in 2026 by human researchers.

One final factor to consider is the pace of development. This is hardly the final form of AI pipeline tools in this category. OpenAI has been running similar tests of the Linux kernel with its frontier models for some time now, but none thus far has been nearly as impressive as the most recent run. The previous test was only several months ago, in July 2026 with the then-leading GPT‑5.6 Sol. That test cost about half as much at $513,690 in API credits, yet only uncovered 166 kernel vulnerabilities. A test in 2025 with an older model turned up only 24.