RefluXFS CVE-2026-64600: patch your Linux servers
A nine-year-old XFS race condition gives any local Linux user root on default RHEL-family installs. Qualys found it with AI help. Patch and reboot.
CVE-2026-64600 — call it RefluXFS — is the kind of bug that changes what a small foothold is worth. It's a race condition in the Linux kernel's XFS filesystem that lets any unprivileged local user overwrite root-owned files and become root. It's been in every mainline kernel since 4.11 shipped in 2017, it's on by default across the RHEL family, and Qualys estimates it touches more than 16.4 million systems. Patch and reboot your Linux boxes.
What actually happened
Per the Qualys Threat Research Unit advisory, the flaw lives in XFS's reflink copy-on-write path. Two concurrent O_DIRECT writes to the same reflinked file cause the kernel to drop its inode lock while waiting on log space. A second writer completes its CoW cycle in that window; the first writer reacquires the lock holding a stale block address, assumes the block is private, and writes straight to the original file's physical location.
In practice: an ordinary user overwrites /etc/passwd or a SUID-root binary. The change survives reboot and produces no kernel log entry. Qualys reports a proof-of-concept that reliably lands host root.
Vulnerable by default means reflink-enabled XFS, which covers RHEL 8/9/10, CentOS Stream, Oracle Linux, Rocky, AlmaLinux, CloudLinux, Fedora Server 31 and later, and Amazon Linux 2023. The fix merged into the kernel tree on July 16 and distributions are backporting it now. SELinux, kernel lockdown, and KASLR do not stop it.
There's a detail worth sitting with. Qualys says it worked with Anthropic and tasked Claude Mythos Preview with hunting for a Dirty COW–style race condition; the model found the XFS race and generated a working proof-of-concept, which human researchers then verified independently before coordinating disclosure. A nine-year-old kernel bug that survived a decade of human review did not survive a model pointed at the right question.
Why local privilege escalation matters for your business
The reflex on local-only bugs is to skip them — nobody has a shell on our server. That reasoning stopped holding sometime last year. Every web app with a file upload, every CI runner executing a pull request, every container escape, every leaked deploy key, every AI coding agent with a sandbox is a path to "unprivileged local user." RefluXFS converts all of them into root, on the same box, with no logs.
The work is short. Run your distro's update and reboot — a live kernel patch that never gets rebooted is a spreadsheet entry, not a fix. Confirm the kernel version afterward rather than trusting the package manager's word for it. Then go find the Linux hosts nobody owns: the old app server, the appliance an MSP dropped in, the VM someone stood up for one project in 2023. Those are where a 2017 kernel is still running.
And note what the discovery story implies about your patch window. If a model can be aimed at a class of bug and produce working exploit code, the interval between "flaw exists" and "flaw is weaponized" is compressing on both sides of the fence. Monthly patching cadences were built for a slower world.
Key takeaways
- CVE-2026-64600 is an XFS reflink race condition giving any local user root; present in every kernel since 4.11 (2017)
- Default-vulnerable on RHEL 8/9/10, CentOS Stream, Oracle Linux, Rocky, AlmaLinux, CloudLinux, Fedora Server 31+, Amazon Linux 2023 — Qualys estimates 16.4M+ systems
- Exploitation overwrites root-owned files, persists across reboot, and writes nothing to the kernel log; SELinux and KASLR don't help
- Fixed in the kernel tree July 16 and backporting now — update and actually reboot, then verify the running kernel
- Qualys found it by aiming Claude Mythos Preview at Dirty COW–style races; assume exploit development is getting faster
Don't know how many Linux hosts you have? That's the real finding here, and it's the one we fix first. We inventory what's running, what kernel it's on, and who's responsible for patching it — then set up a cadence that doesn't depend on anyone remembering. Book an infrastructure audit or see how we handle patching for small teams.
Sources: Qualys TRU, oss-security, BleepingComputer.
- #cve-2026-64600
- #linux
- #privilege-escalation
- #patch-management
- #ai-security
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Qualcomm's price increase hits your hardware budget
Qualcomm told customers of a double-digit price increase on chips shipped after September 1. The AI buildout is now repricing hardware that has nothing to do with AI.
Read it24,000 exposed BMCs leak hashes: close your IPMI port
A 2004 protocol flaw with no patch is handing out password hashes from 24,000 internet-exposed server BMCs. The fix is network exposure, not a firmware update.
Read it