On September 30, 2026, NVIDIA’s September driver bulletin disclosed CVE-2026-47591 and CVE-2026-47501 (both CVSS 7.8) in the kernel layer of its Linux GPU display driver. On November 12, Phoronix published two NVIDIA Linux driver flaws it calls NvRoot and ForgePtr and said 25 more are coming. The first CVE lets an unprivileged user write to memory the driver should keep read-only. The second is an out-of-bounds write during event buffer setup. Both need only a low-privileged local account, so on a shared GPU server every user with a shell is in scope.
This guide gives you the fixed release for each driver branch and shows which version checks you can trust. It then walks through the update on Ubuntu, Rocky Linux and Fedora. Everything was run in November 2026 on Ubuntu 26.04, Ubuntu 24.04 and Rocky Linux 10 with SELinux enforcing. None of the test VMs had a GPU, so the driver built but never loaded, and that gap shapes how you should read the checks.
The two driver bugs and who is exposed
Both CVEs sit in the kernel module, not in CUDA or the userspace libraries. Any process that can open the NVIDIA device nodes reaches that code. NVIDIA’s GPU display driver security bulletin lists the same impact for both bugs: code execution, privilege escalation, data tampering, denial of service and information disclosure.
| CVE | What the bulletin describes | Weakness | CVSS 3.1 |
|---|---|---|---|
| CVE-2026-47591 | Unprivileged user bypasses read-only memory protection and writes to read-only memory | CWE-863, incorrect authorization | 7.8, AV:L/AC:L/PR:L/UI:N |
| CVE-2026-47501 | Out-of-bounds write from mismatched memory buffers during event buffer setup | CWE-787, out-of-bounds write | 7.8, AV:L/AC:L/PR:L/UI:N |
Treat the nicknames with care. NvRoot and ForgePtr are Phoronix’s names. Neither NVIDIA nor the distro trackers use them, and we could not confirm from a public source that they refer to these two CVEs rather than to fixes still under embargo.
Whatever the names cover, the fixed builds below are the ones NVIDIA lists for these two CVEs, so patch to them now and check the bulletin again as more fixes appear. Red Hat rates CVE-2026-47591 Important and limits it to the proprietary display driver, saying standard open source drivers are not affected. Debian’s tracker also marks NVIDIA’s open GPU kernel modules vulnerable, so the open-module packages need the same update.
Your exposure depends on who can log in, not on whether the box faces the internet. A single-user workstation is low risk. A shared training node, a CI runner that runs untrusted jobs, and a cloud host that rents out GPU slices are all high risk, because a tenant or a job is exactly the low-privileged local user these bugs need. Patch those hosts first.
Fixed driver versions by branch
NVIDIA fixed each supported Linux branch separately. A branch is fixed only at or above the version in this table. The bulletin says every earlier release on that branch is affected and warns that older, unlisted branches may be affected too.
| Driver branch | Fixed in (Linux) | Notes |
|---|---|---|
| R615 | 615.71.09 | GPU display driver |
| R610 | 610.57.04 | GPU display driver |
| R595 | 595.91.07 | Same build as the vGPU 20.2 Linux guest driver |
| R580 | 580.178.04 | Same build as the vGPU 19.6 Linux guest driver |
| vGPU Manager 20.2 / 19.6 | 595.91.04 / 580.178.05 | Host side of NVIDIA vGPU, both CVEs listed |
| R620 | Not in the bulletin | Shipping in the repos; NVIDIA has not stated its status |
| R570, R575, R590 and older | No fix listed | Move to a listed branch |
R620 is the one that will trip people up. NVIDIA’s download server and the CUDA repositories already carry 620.70.06, and package managers pick it as the default. The bulletin, last updated October 31, does not mention R620 at all. A newer build probably carries the fix, but the 620.70.06 README does not say so either. Until NVIDIA lists it, we treat R620 as unconfirmed and hold production hosts on R595 or R580, which the bulletin does cover.
Driver version checks you can trust
Five commands report the driver version, and they don’t all read the same thing. Run them together. The first two read the module the kernel is running. modinfo reads the module file on disk, and the package queries read the package database:
nvidia-smi --query-gpu=driver_version --format=csv,noheader
cat /proc/driver/nvidia/version
modinfo -F version nvidia
dpkg -l | grep -E 'nvidia-(open|driver|dkms|kernel)'
rpm -qa | grep -Ei '^(nvidia|kmod-nvidia|akmod-nvidia|xorg-x11-drv-nvidia)'
Here is that set on the Ubuntu 26.04 lab VM after the update. The module file and the DKMS build both say 595.112.07, which is above the R595 fix, and dpkg lists the same version for every driver package. nvidia-smi and /proc have nothing to report, because no module is loaded:

That split is the trap after any driver update. modinfo and dpkg or rpm report the new version as soon as the package lands, while the old module keeps running in the kernel. The host stays vulnerable until it reboots or the module is reloaded. Only /proc/driver/nvidia/version and nvidia-smi prove what is running:
/proc/driver/nvidia/versionshows the loaded kernel module and is the line to compare against the table.nvidia-smireports the same loaded version and fails when no driver is running.- A “Driver/library version mismatch” error from
nvidia-smimeans the old module is still loaded. modinfo -F version nvidiamatches the loaded version only after a reboot or a module reload.- The package version matches the module file only when the DKMS or akmods build succeeded for the running kernel.
If /proc and modinfo disagree on a GPU host, you have patched the disk but not the kernel. Schedule the reboot. Our nvidia-smi command reference covers more query fields, which help when you check a whole fleet.
Ubuntu archive packages already at the fix level
Ubuntu’s CVE tracker still lists the nvidia-graphics-drivers 580 and 595 sources as “Needs evaluation” on 26.04 and 24.04, and no USN exists. That reads like nothing has shipped. The archive says otherwise. On both releases, apt already offers driver builds at or above the fixed versions:
sudo apt-get update
apt-cache policy nvidia-driver-580-open nvidia-driver-595-open nvidia-driver-580 nvidia-driver-595 2>/dev/null | grep -E '^nvidia|Candidate|Installed'
On Ubuntu 26.04 the 580 metapackages resolve to 580.178.04, exactly the R580 fix, and the 595 metapackages resolve to 595.99.02, above the R595 fix. Ubuntu 24.04 showed the same versions with a 24.04.1 suffix:

If your host runs the Ubuntu-packaged nvidia-driver-580 or nvidia-driver-595 metapackage, a normal apt upgrade moves it to these builds. Then reboot and read /proc/driver/nvidia/version. Older branch names behave differently on each release. On 26.04, 535, 550, 570 and 590 are transitional packages that pull in a current branch. On 24.04, 550, 570, 575 and 590 are still real metapackages, and the bulletin lists no fix for those branches. A 24.04 host on one of them has to switch to nvidia-driver-580 or nvidia-driver-595. Our NVIDIA driver and CUDA install guide for Ubuntu covers the move.
CUDA repository installs and the R620 default
Hosts that take the driver from NVIDIA’s CUDA repository face a different risk. An unpinned upgrade jumps to R620, the branch the bulletin does not cover. Install the repository keyring the usual way (shown here for 26.04; use ubuntu2404 in the path on 24.04):
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2604/x86_64/cuda-keyring_1.1-1_all.deb
sudo dpkg -i cuda-keyring_1.1-1_all.deb
sudo apt-get update
Now look at what nvidia-open would install. The repository carries every branch side by side, from 595.58.03 up to 620.70.06, and the candidate is the newest:

NVIDIA ships a pin for this case. Each nvidia-driver-pinning-<branch> package drops an APT preferences file that raises one branch to priority 1000. The packages conflict with each other, so only one pin can be active at a time. We pinned R595 because the bulletin covers it and the repository has a fixed build on it:
sudo apt-get install -y nvidia-driver-pinning-595
cat /etc/apt/preferences.d/nvidia-driver-pin
The first stanza of the file covers the driver and kernel module sources. The full file also pins nvidia-modprobe, persistenced, settings, the meta packages and Fabric Manager the same way:
# Basic driver packages, includes foreign architectures
# cuda-compat, built from nvidia-graphics-drivers does not have a generic common name so APT installs it
# if requested.
Package: src:nvidia-graphics-drivers:any src:nvidia-kmod-open:any
Pin: version 595*
Pin-Priority: 1000
Check the pin before you trust it. apt-cache policy nvidia-open now gives the 595 builds priority 1000 and makes 595.112.07 the candidate, while 610, 615 and 620 stay at the default 600:

With the pin in place, install or upgrade the driver. The DKMS build is the slow part. It took 4m18s on the Ubuntu 26.04 VM and 3m07s on 24.04, so plan for a few minutes per host in a maintenance window:
sudo apt-get install -y nvidia-open
On 24.04 the repository still carries the R580 branch, with nvidia-driver-pinning-580 and pins for exact builds such as nvidia-driver-pinning-580.178.04. The 26.04 repository starts at R595. If your workloads have to stay on R580, use one of those pins and confirm the candidate is 580.178.04 or later. Run sudo dkms status afterwards and confirm the module built for your running kernel before you reboot. On a host with Secure Boot on, the rebuilt module also has to be signed with a key enrolled in MOK, or it will not load after the reboot and the GPU stays down.
Rocky Linux and RHEL through the CUDA repo
On Rocky Linux 10 and RHEL 10 the CUDA repository has the same problem: unpinned, dnf takes R620. It has no pinning package for EL. Instead, each nvidia-open build provides a branch capability, nvidia-open-595, and you can install against that to stay on the branch. Add the repository and the headers DKMS needs:
sudo dnf config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel10/x86_64/cuda-rhel10.repo
sudo dnf -y install epel-release
sudo dnf -y install kernel-devel-$(uname -r) kernel-headers
That block is the Rocky Linux path we ran. RHEL does not ship epel-release in its own repositories. NVIDIA’s RHEL steps enable the CodeReady Builder repository with subscription-manager and install EPEL from the Fedora project’s epel-release-latest-10 RPM. Also, kernel-devel-$(uname -r) only resolves while the running kernel is still the newest in the mirrors. If it fails, update and reboot into the current kernel first.
Then install the branch rather than the package. The install pulled the whole open driver stack plus a desktop dependency tree, and it took 5m14s on the lab VM:
sudo dnf -y install nvidia-open-595
The provides list explains why this works. The 595.112.07 build announces nvidia-open-595, so dnf picks the newest R595 build and never crosses into 610 or 620. The installed packages and the DKMS build agree, and modprobe fails only because the VM has no GPU:

A later dnf upgrade can still pull a newer branch, because the installed nvidia-open package itself has no lock. Lock it with the versionlock plugin. We installed python3-dnf-plugin-versionlock on the lab VM but did not apply a lock. Adapted from NVIDIA’s version-locking page for the 595 branch, the command is sudo dnf versionlock add '*nvidia*595*'. It locks the exact installed build, not the branch, so it also blocks the next R595 security fix. Run sudo dnf versionlock delete '*nvidia*595*' before you take that fix. The branch-capability install is our approach, not NVIDIA’s documented one, and an alternative to the lock is to keep the branch name in every automation run. SELinux stayed enforcing through the install and the DKMS build, so none of this needs a policy change.
Fedora and RPM Fusion akmod builds
Most Fedora GPU users run RPM Fusion’s akmod-nvidia, and the risk there is a stale release repository. We queried Fedora 44 and 45 in clean containers on the Rocky VM. Fedora 44’s base nonfree repository still ships 595.58.03, which is below the R595 fix:

Fedora 45 started from a newer base. Its release repository carries 615.78.08, which is already past the R615 fix, and the updates repository has R620 as on 44:
Fedora release 45 (Forty Five)
Available packages
akmod-nvidia.x86_64 3:615.78.08-1.fc45 rpmfusion-nonfree
akmod-nvidia.x86_64 3:620.70.06-1.fc45 rpmfusion-nonfree-updates
xorg-x11-drv-nvidia.x86_64 3:615.78.08-1.fc45 rpmfusion-nonfree
xorg-x11-drv-nvidia.x86_64 3:620.70.06-1.fc45 rpmfusion-nonfree-updates
RPM Fusion has no fixed R595 build for Fedora 44, and the main akmod-nvidia package only moves forward to R620, which is unconfirmed. Our container query missed one package: RPM Fusion’s repository index also lists a separate akmod-nvidia-580xx build at 580.196.07 in Fedora 44 updates and in the Fedora 45 release repository. That is above the R580 fix and on a branch the bulletin covers. To stay on a confirmed branch, swap akmod-nvidia and xorg-x11-drv-nvidia for their 580xx packages. They replace the main stack rather than install beside it, and we only listed them, we did not run the swap. To take the newest branch, update to R620.
Fedora 43 is supported until December 2, and its updates repository carries akmod-nvidia 580.178.04, exactly the R580 fix. Whichever route you pick, make sure rpmfusion-nonfree-updates is enabled, upgrade, and give akmods time to finish the module build before you reboot. A machine installed without updates and never upgraded is still on the vulnerable build. If you set up RPM Fusion recently, our RPM Fusion setup guide for Fedora shows how the repositories are enabled, and the Fedora NVIDIA driver guide covers the akmod workflow from scratch.
Debian, containers and Kubernetes GPU nodes
Debian has no fix yet. Its tracker marks both CVEs vulnerable in bookworm (535 branch), trixie and sid (550 branch), along with nvidia-open-gpu-kernel-modules. Those are non-free packages, which Debian does not fully cover with security support, so don’t wait for a DSA. Debian GPU hosts that need a fix now have to move to NVIDIA’s packaging or the CUDA repository.
For a host that has a GPU it does not use, Red Hat’s mitigation is to stop the module from loading. Add install nvidia /bin/true to a file in /etc/modprobe.d/, then reboot. A plain rmmod nvidia usually fails while nvidia_drm, nvidia_modeset and nvidia_uvm hold the module, and unloading it stops every GPU workload on the host.
Containers do not change any of this. The NVIDIA Container Toolkit gives a container the host’s device nodes and driver libraries, and every container on the node talks to the single kernel module the host loaded. Updating a CUDA base image does nothing for these CVEs. The fix is on the host, in one of two places:
- Nodes with a host-installed driver are patched with the distro steps above. The GPU Operator does not manage drivers installed on the host.
- Nodes running the GPU Operator’s driver container are patched by changing
spec.driver.versionin the ClusterPolicy, orspec.versionin an NVIDIADriver resource, to a fixed build. - Each node’s
nvidia.com/gpu-driver-upgrade-statelabel showsupgrade-doneonce the Operator’s upgrade controller has cordoned it, evicted the GPU pods and reloaded the module.
NVIDIA’s GPU Operator driver upgrade guide covers parallelism and drain settings for that controller. Pick a fixed version that exists as a driver container image for your node OS before you patch. With vGPU, patch both sides. The host vGPU Manager row lists both CVEs, and the Linux guest driver row lists CVE-2026-47501. Each has its own fixed release in the table. If you already export driver metrics, the DCGM exporter setup reports each node’s driver version, which makes fleet tracking easier.
Audit after the reboot
An update is not done until the running kernel proves it. Work through these on each GPU host after its reboot or module reload:
cat /proc/driver/nvidia/versionshows a version at or above the fixed release for its branch.nvidia-smi --query-gpu=driver_version --format=csv,noheaderprints the same version as/proc.modinfo -F version nvidiamatches the loaded version, so nothing newer is waiting on disk.dkms statusor the akmods log shows the module built for the kernel the host booted.- The branch pin is in place:
/etc/apt/preferences.d/nvidia-driver-pinon Ubuntu, a versionlock or branch-named install on EL. - Kubernetes GPU nodes all report
upgrade-done, and the driver pod on each node runs the fixed version. - vGPU hosts run a fixed vGPU Manager, and their Linux guests run a fixed guest driver.
Check the bulletin again before you close the ticket. Phoronix says 25 more NVIDIA driver vulnerabilities are on the way, and NVIDIA revises this bulletin as branches gain fixes. We will update this page when R620 is listed or Ubuntu publishes a USN.