AI Uncovers Long-Hidden Linux Vulnerabilities

At the [un]prompted AI security conference, Nicholas Carlini from Anthropic revealed how he used Claude Code, an advanced AI language model, to identify multiple remotely exploitable vulnerabilities in the Linux kernel, including one that had remained hidden for 23 years. With minimal guidance, Claude Code was able to pinpoint these vulnerabilities by analyzing the Linux kernel's source code. Carlini shared a simple script that iterates over all files in the source tree, instructing Claude Code to identify potential security flaws, simulating a capture-the-flag cybersecurity challenge.

One significant vulnerability was found in the Linux NFS driver, which could allow an attacker to read sensitive kernel memory over the network. This flaw involved a complex understanding of the NFS protocol, demonstrating Claude Code's capability to detect non-obvious bugs. The attack scenario involved two NFS clients interacting with an NFS server, exploiting a buffer overflow when the server attempted to deny a lock request with an insufficiently sized buffer.

This vulnerability had been part of the Linux kernel since 2003, introduced in a patch that predated the release of Git. Carlini has discovered numerous other potential bugs in the Linux kernel but faces a bottleneck in manually validating these findings before reporting them to maintainers. Despite this, he has already reported and fixed several vulnerabilities.

Carlini emphasized the rapid advancement of AI models in uncovering security vulnerabilities, noting that newer versions of Claude Code, like Opus 4.6, have significantly outperformed older models in identifying these issues. He anticipates a surge in discovered security bugs as more researchers and attackers leverage these powerful AI tools.