Linux Ceph CVE-2026-99304: Who Is Exposed, How to Fix
TL;DR · the short version
CVE-2026-99304 is an integer overflow in the Linux kernel Ceph file system client: the bounds check on snapshot data sent by a Ceph metadata server (MDS) can wrap to zero, so the kernel reads past the end of the message buffer. The kernel CVE team published it on 16 November 2026 with its own CVSS 3.1 score of 9.8; Red Hat rates it 7.5, denial of service. The precondition is the part that matters: the attacker has to control or impersonate the MDS a host mounts CephFS from. Fixed upstream in 6.12.117, 6.18.63, 7.3.4 and 7.4-rc1; the 6.6, 6.1, 5.15 and 5.10 longterm branches have no fix yet. Ubuntu and Red Hat had not shipped fixed kernels on 19 November; Debian trixie has. Not in KEV, no reported exploitation. Hosts that never load the ceph module are not reachable.
CVE-2026-99304 is a flaw in the Linux kernel's Ceph file system client. When a Ceph metadata server (MDS) sends snapshot information, the kernel checks that the message is long enough, but two 32-bit counts taken from the network are added together first and the sum can wrap to zero, so the check always passes and the kernel reads past the end of the message buffer. The CVE record, published by the kernel CVE team on 16 November 2026, says a malicious or compromised MDS can trigger it with no action on the client, and gives it its own CVSS 3.1 score of 9.8 (Critical). It is fixed upstream; most distributions had not shipped a fixed kernel when we checked on 19 November.
Who this is for: teams that run Ceph storage, platform teams whose Kubernetes or OpenStack hosts mount CephFS, and anyone triaging a "9.8 critical Linux kernel" alert from a scanner this week.
Based on the CVE record, the kernel.org commits and stable tags, NVD, CISA KEV, FIRST EPSS and the Ubuntu, Debian and Red Hat trackers, all read on 19 November 2026. We could not open the GitHub advisory GHSA-gwm4-f7gj-p7vj, so nothing here relies on it. PlayCISO is not affiliated with the Linux kernel project, Ceph or any distribution.
Confirmed vs not yet known (19 Nov 2026)
- Confirmed by the kernel CVE record and commits: integer overflow in
ceph_update_snap_trace()(fs/ceph/snap.c) leading to an out-of-bounds read; reachable from an unsolicited snapshot message, an MDS reply or a capability import; present since kernel 2.6.34; fixed in 6.12.117, 6.18.63, 7.3.4 and 7.4-rc1. - Who scored it: the 9.8 is the Linux kernel CNA's own score. There is no CISA ADP score; NVD shows the CNA score and has not done its own analysis. Red Hat rates it 7.5 (Important), availability only.
- Not reported: exploitation in the wild. Not in CISA KEV (catalog 2026.11.18). EPSS 0.00157 on 19 November.
- Not fixed yet: the 6.6, 6.1, 5.15 and 5.10 longterm kernel branches; Ubuntu (all releases); Red Hat Enterprise Linux 7 to 10; Debian 12 (bookworm).
- Not shown: code execution. The kernel CNA's scoring notes describe a possible follow-on memory write; no source shows a working exploit.
What the bug is, in plain words
A CephFS client keeps a session open with the cluster's metadata server. The MDS tells the client about snapshots in a compact binary format: a header, then two lists whose lengths are given as 32-bit numbers. Before reading the lists, the client works out how many bytes they take and checks the message is that long. The CVE record explains that the two counts were added as 32-bit numbers before being multiplied, so a sender can pick values whose sum wraps around to zero. The length check then asks for zero bytes, passes, and the kernel copies data from beyond the end of the buffer.
The fix computes the length with the kernel's saturating helpers (size_add() and size_mul()), so a huge count now fails the check instead of wrapping. The same style was already used elsewhere in the Ceph client.
The part that decides your risk is who can send that message: the MDS of a cluster the host has mounted. The CVE record calls it "a malicious or compromised MDS". Red Hat's statement says an attacker must be "controlling or impersonating" a Ceph MDS, and that RHEL systems are vulnerable "only if they mount Ceph file systems and communicate with an untrusted or compromised Ceph storage cluster."
A 9.8 with a real precondition
The CVSS vector (network, low complexity, no privileges, no user interaction, high impact on all three) reads as "anyone on the internet can take over the box". That is not what the record describes. Who scored it matters here:
| Who | Score | What it says |
|---|---|---|
| Linux kernel CNA (CVE record, 16 Nov) | CVSS 3.1 9.8, AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H | Treats the MDS peer as the attacker. Its scoring notes say confidentiality is high (memory read into the snapshot list), integrity high (a later update can write past a short buffer) and availability high (the client fences the mount and needs a remount). |
| CISA ADP | None | No ADP container in the record on 19 Nov. |
| NVD | Not analysed | Status "Received"; displays only the CNA score. |
| Red Hat | CVSS 3.1 7.5 (draft), Important | Denial of service only (A:H). Notes that Ceph traffic normally runs on dedicated, trusted networks. |
| Ubuntu | Priority High | No separate CVSS published on its page. |
Our reading: if you do not mount CephFS with the kernel client, this is not reachable on that host whatever the score says. If you do, the real question is how much you trust the Ceph cluster and the network path to it. A shared or third-party Ceph cluster, or one already compromised, makes this a real way to crash or disrupt every client node at once. A score of 9.8 on the scanner report should not jump this ahead of an exploited, internet-facing flaw.
Useful next, if this was relevant to you
Who is exposed and who is not
- Exposed: Linux hosts with a CephFS mount through the kernel client (a mount of type
ceph). Common cases: Kubernetes nodes where CephFS volumes from Rook or ceph-csi are mounted with the kernel mounter; OpenStack instances using Manila CephFS shares through the kernel client; HPC and file servers that mount CephFS directly. - Not reachable: hosts that never load the
cephfile system module. Most servers and desktops fall here. - Not covered by this CVE:
ceph-fuseand libcephfs, which are userspace clients shipped with Ceph and do not run the kernel's fs/ceph code. No source we read says whether they have a similar flaw. - RBD only: RBD block devices use the kernel's
libcephandrbdmodules, not thecephfile system module, so RBD alone does not reach this function. See CVE-2026-99296 below, which does affect libceph. - The Ceph servers themselves: the bug is in the client. Patching the cluster does not fix the clients.
Affected and fixed versions
The record says the flaw has been present since kernel 2.6.34, when Ceph snapshot support was merged. Upstream fixes, from the CVE record and checked against each stable tag:
| Kernel branch | First fixed release | Commit |
|---|---|---|
| Mainline | 7.4-rc1 (1 Nov 2026) | b3bf89767e3c |
| 7.3 stable | 7.3.4 (10 Nov 2026) | 25ae51760b75 |
| 6.18 longterm | 6.18.63 (10 Nov 2026) | f0dbd6b9b05e |
| 6.12 longterm | 6.12.117 (10 Nov 2026) | bf8f13872959 |
| 6.6, 6.1, 5.15, 5.10 longterm | No fix yet | Latest releases 6.6.162, 6.1.192, 5.15.225 and 5.10.274 still carry the old check (checked 19 Nov) |
Distribution kernels are patched on their own schedule. Status on 19 November 2026:
| Distribution | Status | Source |
|---|---|---|
| Ubuntu (every release it lists, including 24.04 LTS on 6.8 and the 22.04 HWE 6.8 kernel) | "Needed": no fixed package yet. Priority High. | Ubuntu |
| Debian 13 trixie | Fixed in trixie-security 6.12.117-1 | Debian |
| Debian forky / sid | Fixed (7.3.4-1 / 7.3.5-1) | Debian |
| Debian 12 bookworm | Vulnerable (6.1 kernel) | Debian |
| Red Hat Enterprise Linux 7, 8, 9, 10 (kernel and kernel-rt) | Affected, no fix yet | Red Hat |
| OpenShift (RHEL CoreOS 8, 9, 10 images) | Affected, no fix yet | Red Hat |
| Red Hat Enterprise Linux 6 | Not affected | Red Hat |
How to check a host
- Is the module loaded?
lsmod | grep -w ceph. No line means the CephFS client is not loaded on that host right now. (libcephon its own, for example withrbd, is not this code.) - Is anything mounted?
findmnt -t cephorgrep ' ceph ' /proc/mounts. On Kubernetes, look for CephFS persistent volumes and which mounter the CSI driver uses. - Which kernel?
uname -r, then compare with your distribution's tracker, not with the upstream numbers: distribution kernels carry backports under their own version strings. - Could it load later? A host with no mount today can still load the module on demand when a CephFS volume is scheduled to it. Inventory by role (storage-attached node pools), not only by today's
lsmod.
Timeline
| Date | Event | Source |
|---|---|---|
| Kernel 2.6.34 | Ceph snapshot management merged; the CVE record marks 2.6.34 as the first affected version | CVE record |
| 1 Nov 2026 | Mainline 7.4-rc1 tagged with the fix | kernel.org |
| 10 Nov 2026 | Stable releases 6.12.117, 6.18.63 and 7.3.4 tagged with the backport | kernel.org |
| 16 Nov 2026 | CVE-2026-99304 published by the Linux kernel CNA, CVSS 9.8; Red Hat publishes its assessment (Important, 7.5) | CVE record; Red Hat |
| 18 Nov 2026 | CISA KEV catalog 2026.11.18 does not list it (checked 19 Nov) | CISA |
| 19 Nov 2026 | NVD status "Received"; EPSS 0.00157; Ubuntu "needed" on all releases; Debian trixie fixed | NVD; FIRST; Ubuntu; Debian |
Two related Ceph CVEs from the same batch
Among the other kernel CVEs published on 16 November, two more are in the Ceph client code, and are fixed by the same stable branches:
- CVE-2026-99306: the CephFS client accepts an MDS map that claims more ranks than it can hold, causing an out-of-bounds read; the map arrives from the Ceph monitor. Affected since 5.15, CNA score 9.8, fixed in 6.12.117, 6.18.63 and 7.3.4.
- CVE-2026-99296: in
libceph, shared by CephFS and RBD, a short monitor command reply is decoded using stale buffer bytes; the record shows it crashing the kernel throughrbd. Affected since 4.9, CNA score 7.1, fixed in 6.12.117, 6.18.61 and 7.3.2. This one reaches RBD-only hosts.
Patch them together: the same kernel update covers all three on the fixed branches.
What security leaders should do this week
- Find the CephFS clients. List every host and node pool that mounts CephFS with the kernel client. That is your exposed set; everything else can follow the normal patch cycle.
- Re-rank the scanner finding. Scanners will flag every Linux kernel at 9.8. Mark hosts without the ceph module as not reachable and record why, so the finding does not crowd out exploited flaws.
- Patch exposed clients when your distribution ships. Ubuntu and RHEL had no fixed kernel on 19 November; subscribe to their notices and schedule a drain-and-reboot for storage-attached nodes. On upstream kernels, move to 6.12.117, 6.18.63, 7.3.4 or later; 6.6 and older longterm branches have no fix yet.
- Look at trust in the storage path. The attacker has to be the MDS. Keep Ceph traffic on a dedicated network, keep cephx authentication on, restrict who administers the cluster, and treat a shared or third-party CephFS service as a higher-risk dependency.
- If you cannot patch soon, reduce the module's reach. On hosts that should never mount CephFS, blocking the ceph module from loading removes the exposure entirely. Do not do this on hosts that need CephFS.
- Record the decision. Not in KEV, no reported exploitation, a hard precondition, and a high vendor score. Set a deadline you can defend and log any exception with an owner.
Put it to work
- SSVC calculator: turn exploitation, exposure and mission impact into a patch decision that does not depend on the 9.8.
- EPSS lookup: watch whether the exploitation probability for CVE-2026-99304 moves.
- KEV checker: re-check whether it has been added to CISA's exploited list.
- Vulnerability prioritizer: rank it against the rest of your kernel and storage backlog.
The vendor timeline is in the Breach Index entry for the Linux kernel. It is a vulnerability record, not a breach.
FAQ
What is CVE-2026-99304?
CVE-2026-99304 is a flaw in the Linux kernel's Ceph file system client (fs/ceph). When the client checks the size of the snapshot data a Ceph metadata server (MDS) sends, two 32-bit counts from the network are added before the check and the sum can wrap to zero. The check then passes and the kernel reads past the end of the message buffer. The kernel CVE team says a malicious or compromised MDS can trigger it with no action on the client.
Why is CVE-2026-99304 rated 9.8?
The 9.8 (CVSS 3.1, network, no privileges, no user interaction) is the Linux kernel CNA's own score in the CVE record published on 16 November 2026. There is no CISA ADP score, and NVD has not analysed the CVE: it shows only the CNA's score. Red Hat rates the same flaw Important, 7.5, availability impact only, and says systems are vulnerable only if they mount Ceph file systems and talk to an untrusted or compromised Ceph cluster.
Who is exposed to CVE-2026-99304?
Linux hosts that mount CephFS with the kernel client: for example Kubernetes nodes where a CephFS volume (Rook, ceph-csi) is mounted with the kernel mounter, OpenStack instances using Manila CephFS shares through the kernel client, and servers with a mount of type ceph. The attacker has to control or impersonate the MDS the client talks to. A host that never loads the ceph file system module cannot reach this code.
Does CVE-2026-99304 affect ceph-fuse or RBD?
The CVE record names only the kernel file fs/ceph/snap.c. ceph-fuse and libcephfs are userspace clients that ship with Ceph itself and do not run this kernel code, so this CVE does not cover them; no source we read says whether they have a similar flaw. RBD block devices use the kernel's shared libceph and rbd modules, not the ceph file system module, so RBD alone does not reach this function. A separate CVE posted the same day, CVE-2026-99296, is in libceph and does reach rbd.
Which kernel versions fix CVE-2026-99304?
Per the CVE record and the stable tags: 6.12.117, 6.18.63 and 7.3.4 (all tagged 10 November 2026), and mainline from 7.4-rc1. The record lists no fixed release for the 6.6, 6.1, 5.15 or 5.10 longterm branches, and their latest releases on 19 November 2026 (6.6.162, 6.1.192, 5.15.225, 5.10.274) still have the old check. Distribution kernels are patched separately: check your distribution's tracker.
Is there an Ubuntu or Red Hat fix for CVE-2026-99304?
Not yet, as of 19 November 2026. Ubuntu lists the linux package as "needed" on every supported release, including 24.04 LTS (the 6.8 kernel) and the 6.8 HWE kernel on 22.04, with priority High. Red Hat lists the kernel on RHEL 7, 8, 9 and 10 and the RHEL CoreOS images in OpenShift as affected, with no fix yet; RHEL 6 is not affected. Debian has fixed it in trixie-security (6.12.117-1), forky and sid; bookworm is still listed as vulnerable.
Is CVE-2026-99304 being exploited?
We found no report of exploitation. It was not in the CISA Known Exploited Vulnerabilities catalog (version 2026.11.18) when we checked on 19 November 2026, and its EPSS score that day was 0.00157 (about 0.16% predicted chance of exploitation in the next 30 days).
Can CVE-2026-99304 lead to code execution?
No source we read shows code execution. The fix commit describes an out-of-bounds read. The kernel CNA's scoring notes go further and describe a follow-on path where a later snapshot update writes past a short buffer, which is why it rates integrity High; Red Hat's assessment treats it as denial of service. Treat it as a reliable way for a hostile MDS to crash or disrupt a client, with memory disclosure and corruption possible.
Sources
- CVE record: CVE-2026-99304 (CNA: Linux, published 16 Nov 2026). Primary source for the description, affected range, fixed versions and the CNA's CVSS 9.8 and scoring notes.
- kernel.org: ceph: fix integer overflow in the snap trace bounds check (mainline b3bf89767e3c), with stable backports bf8f13872959 (6.12), f0dbd6b9b05e (6.18) and 25ae51760b75 (7.3).
- NVD: CVE-2026-99304 (status Received, checked 19 Nov 2026)
- Red Hat: CVE-2026-99304 and Red Hat Bugzilla 2573829
- Ubuntu: CVE-2026-99304
- Debian security tracker: CVE-2026-99304
- CISA Known Exploited Vulnerabilities catalog (version 2026.11.18, checked 19 Nov 2026: not listed)
- FIRST EPSS API (score for 19 Nov 2026)
- CVE record: CVE-2026-99306 and CVE record: CVE-2026-99296
This article is educational guidance for security and infrastructure leaders. Distribution fix status changes quickly: check your vendor's tracker before acting.