Last week, OpenAI announced it had discovered 948 Linux kernel vulnerabilities and generated patches it offered to the upstream Kernel.org project. It’s a real milestone for AI in security, and it’s what many of us hoped AI would do.
At CIQ, we see it as the start of the work. Our kernel engineers use AI to explore fixes, and our experience is consistent: only a fraction of AI-generated Linux vulnerability patches can ship without modification, even with the best models available. Finding a bug and drafting a patch are now fast. Shipping a fix that’s safe, maintainable, and compatible across years of supported kernels still takes expert engineers.
OpenAI’s own testing points the same way. Patches that fixed the exploit could still introduce performance issues, violate coding conventions, or miss variants of the vulnerability.
Ask four questions before a patch ships
Kernel maintainers, including CIQ’s kernel team, weigh four questions before accepting any patch, whether a human or an AI wrote it:
- Does it actually fix the vulnerability? A CVE and a CVE with a proof-of-concept exploit are vastly different. Without a reliable way to reproduce the vulnerability, proving a patch addresses it is difficult.
- Does it introduce regressions? Fixing one vulnerability can create new bugs, break existing behavior, or even open a new vulnerability. When a patch touches sensitive components, the medicine can be worse than the poison.
- Is it maintainable? Code quality matters in a project as large and critical as the kernel. LLMs can produce brute-force fixes that work in isolation yet ignore the broader context of the codebase.
- Does it affect compatibility? Linus Torvalds lives by the rule “we do not break userspace.” When a bug has existed for years, software gets built around that behavior, and fixing it can break applications.
No automated test suite can reliably validate every consequence of a kernel patch. That’s why every AI-generated patch still needs human review.
Count the cost of review
Upstream maintainers could soon receive hundreds of AI-generated patches, each needing intense scrutiny. If those patches take more review than they save, the work moves from writing code to reviewing it, and maintainers carry the load.
For enterprise Linux, acceptance upstream is only the first step. Vendors like CIQ then backport fixes to older supported kernels and validate compatibility before anything reaches customers.
Hold humans to the same standard
Humans make mistakes too. People still write most of the code in use today and most vulnerability patches, and they can miss part of a vulnerability, introduce regressions, or leave code that’s painful to maintain.
So start by deciding how to measure success. Today, comparing AI patches to human patches is a tradeoff: AI wins on velocity and cost, and experienced engineers win on accuracy and maintainability. For enterprise Linux, maintainability is mandatory.
Pair AI speed with human judgment
AI-generated patches can be good, and they save real time exploring options. At CIQ, an engineer reviews every one.
Engineers bring three things LLMs still lack: judgment, accountability, and context. An engineer’s name goes on every open source commit, so a bug found later reflects on them professionally and personally. That creates a strong incentive to fix the problem, avoid new issues, and leave an elegant approach for the next engineer. An LLM focuses only on the problem in front of it, which could mean 2,000 lines of code when 30 would do.
We saw this play out recently. When CVE-2026-98432 turned critical after Anthropic demonstrated a working exploit, our kernel team patched every affected kernel stream within three days. The upstream fix was small. The time went into confirming it fixed the vulnerability, introduced no regressions, and would stay maintainable across more than a dozen kernel streams.
We believe AI will get better at producing patches that ship without changes. As an industry, we’re still on that journey.
Watch the benchmark
CIQ will track how many of OpenAI’s patches are accepted upstream, how many need modification, and how much review it takes to get there. The results will offer a useful benchmark for AI-generated vulnerability patches. We’ll share an update next month, so subscribe to our newsletter.
