New RefluXFS Linux flaw lets attackers gain root privileges

RefluXFS: A Nine-Year-Old Linux Kernel Vulnerability

Qualys Threat Research Unit (TRU) discovered RefluXFS (CVE-2026-64600), a race condition that allows local attackers to overwrite protected files and gain root privileges. A patch, commit 2f4acd0, was subsequently added to the Linux kernel source tree on July 16.

This isn't a new vulnerability; it was introduced in February 2017 with commit 3c68d44a2b49 in kernel version 4.11, meaning it has been present in systems for over nine years, affecting any Linux kernel v4.11 or later running an XFS filesystem with reflink enabled. Reflink is enabled by default on most major enterprise Linux distributions, including Red Hat Enterprise Linux (RHEL), Oracle Linux, Amazon Linux, Fedora, CentOS Stream, Rocky Linux, AlmaLinux, and CloudLinux. Qualys estimates over 16.4 million systems are affected.

How RefluXFS Works: The Race to Root

Let's break down the attack chain:

1. Target Selection: An attacker identifies a critical target file, like /etc/passwd or an SUID-root binary such as /usr/bin/su. The only requirement is write access to a directory on an XFS filesystem.
2. Reflink Clone: The attacker uses the reflink feature to clone the target file into a scratch file they own. This creates a copy-on-write link, meaning both files initially share the same physical data blocks on disk.
3. The Race: The attacker then initiates concurrent O_DIRECT writes to their scratch file. O_DIRECT writes bypass the page cache, writing directly to disk.
4. The Lock-Drop Window: The kernel's copy-on-write allocation path for XFS has a specific lock-drop window. During this brief period, the kernel releases a lock that protects the physical block allocation.
5. The Overwrite: One of the attacker's O_DIRECT writes lands in the physical block that is still backing the original target file, before the copy-on-write mechanism can properly allocate a new, separate block for the scratch file.

The outcome is that the attacker's data directly overwrites the original, protected file on disk. It doesn't touch the target file's inode, so if you're modifying an SUID binary, its SUID bit stays intact. This exploitation aligns with MITRE ATT&CK technique T1068 (Exploitation for Privilege Escalation) and specifically T1548.001 (Abuse of SUID/SGID binaries) when targeting such executables. The change persists across reboots, and importantly, it leaves no kernel log output. Crucially, this makes the attack highly stealthy.

Standard defenses like SELinux, kernel lockdown, container isolation, KASLR, SMEP, and SMAP don't stop RefluXFS. Its operation at the filesystem allocation layer, below many standard protections, explains its effectiveness and the difficulty in discovering it.

The Role of AI in Discovering RefluXFS

The innovative method of RefluXFS's discovery is as notable as the vulnerability itself. Qualys didn't stumble upon this. They were actively hunting for Dirty COW-like race conditions as part of a research initiative with Anthropic, integrating the AI model Claude Mythos Preview into their manual audit workflow.

Manually finding these kinds of race conditions is incredibly challenging, requiring precise timing and specific conditions. The fact that an AI model, even with human guidance, could pinpoint such a complex, long-standing flaw that human researchers missed for nine years is a significant development that highlights a shift in vulnerability research.

Expert analysis indicates a clear consensus: AI is becoming highly proficient at identifying programming flaws. This shifts the conversation about the changing role of human researchers, suggesting AI serves as a powerful efficiency multiplier rather than a replacement. Integrating these tools effectively, with human oversight and strategic direction, is crucial to avoid a significant disadvantage against adversaries already adopting such capabilities.

Immediate Actions and Future Implications

The immediate priority is to patch your kernels. Qualys recommends immediate patching. There are no effective mitigations or temporary configuration changes that can stop RefluXFS. Once you've patched, make sure to reboot your systems to verify the update. This is particularly critical for exposed systems or multi-tenant environments where local users might have access.

Looking ahead, this incident shows us an indicator of future trends. This incident highlights a broader trend: AI's capabilities are expanding beyond text and image generation to become an effective tool for vulnerability research. This means we're likely to see more complex, long-standing flaws unearthed at an increased rate. Our defenses must adapt by integrating AI into security operations, not only for vulnerability discovery but also for threat analysis, attack path prediction, and automated response. Human expertise remains critical, but our role shifts from manual bug hunting to strategic oversight and AI-driven threat intelligence. RefluXFS joins a lineage of significant Linux privilege escalation vulnerabilities discovered in recent years, including CIFSwitch, PinTheft, Copy Fail, Dirty Frag, Fragnesia, Pack2TheRoot, and DirtyDecrypt/DirtyCBC. This pattern underscores the ongoing challenge of securing complex operating system kernels against sophisticated attacks. Effectively addressing this evolving threat landscape will necessitate a strategic integration of AI into security operations, complementing human expertise for vulnerability discovery, threat analysis, and automated response.