Claude Opus 4.6 found a 23-year-old Linux kernel vulnerability in a single weekend

Twenty-three years. And one weekend. That is how long it took a simple Bash script and a language model to find a vulnerability that had sat quietly in the Linux file system code the entire time. Researcher Nicholas Carlini started the script on his laptop on Friday evening. By Sunday morning, he had an exploit that worked on every corporate file server released since March 2003.

The story became one of the most discussed events around the release of Linux 7.0, which shipped on April 12, 2026. The new kernel brought two headline changes: Rust is now officially a full-fledged language for kernel development, and AI appears for the first time in the list of code co-authors. Linus Torvalds called this "the new normal."

How Carlini's script worked

The setup is almost embarrassingly simple. The script loops over the Linux kernel source files. It sends each one to Claude Opus 4.6 with a CTF-style system prompt: "imagine you're hunting for vulnerabilities in this code — what could you break here?" Then it moves on to the next file. And the next. No sophisticated tooling, no static analysis — just repeated questions to the model.

Carlini had been running similar scripts for months. The output was always noise: irrelevant patterns, false positives, obvious findings. Then one evening, the model produced something that stopped him mid-sentence.

The model pointed to the code that implements the network file-sharing protocol. That same code runs on file servers in corporate networks, in hospital storage systems, on school shared drives — and on a significant share of the server infrastructure at AWS, Google Cloud, and Azure.

What the hole made possible

The vulnerability required no exploit chain, no stolen credentials, and no administrator privileges. An attacker only had to be on the same local network as the target server and run a small script. From there, they had full control: reading any file, deleting data, installing a persistent backdoor.

Carlini laid out the scenario bluntly. An intern on their first day, connected to the office guest Wi-Fi, could theoretically access HR's payroll records, delete the accounting archive, and copy backups of executives' email. And the backdoor they installed along the way would survive the server's next three reboots.

No admin password. No stolen credentials. Just a Wi-Fi connection.

The bug lived in the kernel from March 2003 to April 2026 — twenty-three years in one of the most scrutinized repositories in the world.

What this means for kernel security

Carlini's case immediately became an argument for using AI in security audits — not only to generate new code, but also as a tool for systematically analyzing existing codebases. The vulnerability was patched in time, and the fix landed in Linux 7.0.

Rust, as the kernel's second official language, partly addresses the class of problems this story belongs to: memory-management bugs in C code. But after 35 years of development, the Linux kernel has enough legacy C components to keep that problem alive for years to come.

What stands out is something else: the model did nothing that couldn't be described as "carefully reading code while asking what could go wrong here." That is what makes the story both alarming and encouraging. Alarming, because anyone can use the same approach. Encouraging, because until now nobody had applied it systematically. By all indications, that is about to change.