RefluXFS Linux Kernel Vulnerability Lets Attackers Gain Root Access

A new Linux vulnerability dubbed “RefluXFS” is a race condition in the Linux kernel’s XFS filesystem copy-on-write path that lets an ordinary local user silently overwrite protected system files and seize host root privileges, even on systems running SELinux in Enforcing mode.

The vulnerability tracked as CVE-2026-64600 uncovered by Qualys Threat Research Unit (TRU) exploits a timing flaw triggered when two concurrent O_DIRECT writes target the same reflinked file on an XFS volume.

XFS normally handles writes to shared blocks by allocating a new private block and remapping the file, but the kernel briefly drops its inode lock while waiting for transaction log space, creating a race window.

A second writer can slip in during this window, remap the file, and drop the reference count — causing the first writer to mistakenly write directly to the original block using a stale reference. Because O_DIRECT bypasses the page cache and skips revalidation, the corrupted write lands permanently on disk.

RefluXFS Linux Kernel Vulnerability

An unprivileged local user can weaponize this flaw to overwrite any readable file on a reflink-enabled XFS volume at the block layer, converting directly into host root access.

Qualys’s proof-of-concept demonstrated this on a default RHEL 10.2 deployment, silently stripping the root account’s password protection within seconds and granting immediate, passwordless root access. The modification survives reboots and leaves zero kernel log artifacts, making detection extremely difficult.

The bug has quietly existed in every mainline and stable Linux kernel since version 4.11, released in 2017. Qualys estimates the flaw potentially impacts over 16.4 million systems worldwide based on analysis using its CyberSecurity Asset Management platform.

Vulnerable systems must meet three conditions: a kernel version 4.11 or later without the patch, an XFS filesystem with reflink=1 enabled, and a directory writable by an unprivileged user alongside a high-value target like a SUID-root binary.

RefluXFS operates at the filesystem allocation layer, below the reach of conventional kernel hardening. Memory-protection mechanisms like KASLR, SMEP, and SMAP target different attack surfaces entirely, kernel lockdown places no restriction on O_DIRECT or FICLONE calls, and SELinux does not block the vulnerable code path in testing.

Container isolation techniques, including user-namespace restrictions and capability limits, likewise fail to touch this layer, meaning there is currently no reliable workaround short of patching.

Notably, this find emerged from a joint research initiative between Qualys and Anthropic, integrating the Claude Mythos Preview model into TRU’s manual audit workflow under Anthropic’s Project Glasswing.

Researchers tasked the model with hunting for Dirty COW-style race conditions; after iterative prompting, it identified the XFS flaw and generated a working proof-of-concept, which Qualys engineers then independently verified before coordinating disclosure with upstream maintainers.

This pattern of AI-accelerated, human-validated vulnerability research mirrors a broader 2026 trend, following similarly significant Linux privilege-escalation disclosures this year such as CVE-2026-46333 (a nine-year-old ptrace flaw) and CrackArmor, an AppArmor-based root-access chain affecting over 12 million systems.

Vendor-fixed kernels are already available and being backported across enterprise distributions including RHEL, Oracle Linux, AlmaLinux, Rocky Linux, and Fedora.

Organizations should prioritize patching internet-facing and multi-tenant systems immediately, apply the relevant distribution security update, and perform a full reboot to confirm the fix is active since no interim mitigation or configuration workaround currently neutralizes the flaw.

These four vulnerabilities emerged within roughly three months of each other and represent an escalating pattern of researchers finding new ways to trick the kernel into writing attacker-controlled data into protected memory or disk blocks that back setuid binaries or config files.

The Privilege Paths Attackers See That You Don’t: BeyondTrust Pathfinder Platform Does It for You -> Get Free Identity Security Assessment