OpenAI uses AI to find 1,724 Linux kernel vulnerabilities for $1.21 million

OpenAI used the GPT-5.5 and GPT-5.6 Cyber models to scan the Linux kernel's source code and, within hours, identified 2,300 potentially dangerous points. Of these, 1,724 were confirmed as real vulnerabilities, and for about 950 of them the agents went as far as developing functional exploits capable of granting root privileges, the highest level of access on a Linux system. The information was reported by the Russian site CNews on November 6, 2026.

The process was divided into automated stages. Half a day of initial scanning of the kernel code flagged the 2,300 suspicious points. Then came 30 hours of building and testing exploit prototypes, which narrowed the list down to the 1,724 confirmed vulnerabilities. In the final stage, the agents themselves prepared finished versions of the exploits and patch suggestions to fix the flaws found.

The number that stands out: $1,279 per exploit

The total cost of the experiment was $1.21 million. Divided by the number of exploits produced, that comes to about $1,279 per exploit. The Russian outlet OpenNet, cited by CNews, compared that figure to what human researchers charge in security competitions such as Pwn2Own and TyphoonPWN, where a single exploit typically costs between $20,000 and $70,000.

In practice, that puts the AI's cost between 16 and 55 times cheaper than a specialized human researcher, according to the same comparison. It's this cost differential, not just the volume of vulnerabilities, that makes the experiment relevant: kernel-level security audits stop being a luxury reserved for the few companies with budget to hire offensive research teams.

Exploits worked even with modern defenses active

The exploits generated by the agents were tested on the most widely used Linux distributions, with the kernel's main memory protections enabled: SMEP (Supervisor Mode Execution Prevention), SMAP (Supervisor Mode Access Prevention), KASLR (Kernel Address Space Layout Randomization), and HARDENED_USERCOPY. Even with these barriers turned on, some of the exploits worked.

Among those that got past the defenses, 21 allowed privilege escalation and 25 enabled information theft. That means the kernel's standard mitigations, designed precisely to make memory-bug exploitation harder, weren't enough to block the AI-generated code in every case.

Not everything was new, and not everything stayed private

OpenAI passed the findings on to the Linux kernel developer community. Some of the vulnerabilities had already been found by human testers, and for some of them patches were already ready or in the works, according to CNews. In other words, the experiment didn't replace the community's work from scratch: for the most part, it confirmed and sped up findings that were already in progress.

It's still unclear whether information about the flaws found by the AI reached the public domain. The concern isn't idle: in October 2026, experts involved in fixing kernel bugs accidentally leaked data about a series of critical vulnerabilities, according to an earlier CNews report. A leak of that kind, combined with a list of nearly 2,000 newly mapped flaws, is exactly the scenario that critical infrastructure maintainers want to avoid.

Not an isolated case: the trend had been building for months

Using AI to hunt for bugs in the kernel didn't start now. At the launch of Linux 7.2, on August 16, 2026, AI models had already helped identify 1,500 vulnerabilities of varying severity, according to a CNews report from September 2026. Neural networks can scan sections of code that go years without human review, something unfeasible at scale for small teams of volunteer maintainers.

But the volume also has a side cost: at the launch of Linux 7.3, on October 18, 2026, between a third and half of the 648 proposed patches were low-priority fixes, tweaks, or clarifications generated by AI. That increased the triage burden on the community, which has to review each suggestion before accepting the merge.

What changes for those who maintain or contribute to open source projects

For those who develop or maintain software that depends on the Linux kernel, directly or indirectly (and that includes practically any backend stack, container, or cloud infrastructure running on Linux distros), OpenAI's experiment points in two directions that don't cancel each other out:

• Security auditing becomes more accessible. A cost of ~$1,279 per confirmed exploit is a fraction of what human researchers charge, which could make kernel audits and audits of other critical infrastructure components cheaper for companies that currently don't have the budget for it.
• The volume of findings becomes a triage problem. Projects maintained by a handful of volunteers have already reported being overwhelmed by low-priority AI-generated patches, as the Linux 7.3 cycle showed. Receiving nearly 2,000 vulnerability reports at once, even if some are already known, requires a triage process that most simple open source projects don't have.

For smaller teams maintaining libraries or internal services on Linux, the practical lesson is different: code scanners based on models like GPT-5.5 and GPT-5.6 Cyber stop being the exclusive domain of big tech and become something mid-sized security teams can consider running against their own code, not just against the kernel.

What remains open

CNews doesn't detail whether OpenAI published the full list of 1,724 vulnerabilities, nor the kernel community's patching timeline. There's also no confirmation, in the available material, of how many of the 950 root exploits still work today, November 7, 2026, against already-patched versions. These points determine whether the episode turns into a successful case of coordinated responsible disclosure, or whether some of these flaws circulate unpatched before reaching the distributions devs actually run in production.