NVIDIA Linux GPU driver: unprivileged user writes to read-only files and gets host root (CVE-2026-47591 "NvRoot", CVE-2026-47501 "ForgePtr")
Severity: high · Patch status: partial · Since 2026-09-30 · CVE-2026-47501, CVE-2026-47591
Two bugs in NVIDIA's Linux GPU kernel driver let any local user with access to a CUDA-capable GPU become root on the host. The user needs no special group and no CAPSYSADMIN, and this works from inside an unprivileged GPU container. That is exactly what a GPU cloud or shared training cluster hands its tenants. NvRoot (CVE-2026-47591): when the driver pins memory for the GPU, it takes the page permission from the CPU-side field alone. So a request marked "CPU read-only, GPU writable" gets a writable GPU mapping of a read-only file's page cache. Stealthium used this to overwrite the cached copy of /usr/bin/su. The bug is in the shared resource-manager code, so both the open and the proprietary builds are affected. It has been in every open-source driver since 515.43.04 (May 2022). It is fixed in 610.57.04, 595.91.07 (595.44.14 on the 595.44.x line) and 580.178.04, and later releases on those branches. Stealthium says R590, R575, R570 and R565-and-earlier have no fix. ForgePtr (CVE-2026-47501): during event-buffer setup, a caller-supplied handle is turned into a kernel pointer without checking that the conversion worked. Together with a separate address leak, that gives an arbitrary kernel write and root. Stealthium reports 610.57.04 is still vulnerable and names 615.71.09 as the fix. NVIDIA published both in its September 2026 GPU display driver bulletin. Stealthium released the write-ups on 12 November 2026 and withheld proof-of-concept code. It says no sysctl, seccomp profile or cgroup closes either bug, so updating the driver is the only fix.
What to do
- Inventory every Linux host with an NVIDIA GPU and its driver version: `cat /proc/driver/nvidia/version` or `nvidia-smi --query-gpu=driver_version --format=csv,noheader`.
- Patch shared and multi-tenant GPU hosts first (rented GPU slices, shared training clusters, Kubernetes GPU nodes, CI runners with GPUs). Any tenant with a shell or a GPU container can reach these bugs.
- Move to 615.71.09 or later where you can: it is the only version Stealthium names as fixing both bugs. 610.57.04, 595.91.07 (595.44.14 on 595.44.x) and 580.178.04 fix NvRoot only.
- If you run R590, R575, R570 or R565-or-older, no fixed release exists on that branch. Move to a fixed branch rather than waiting.
- Don't wait for distro packages: as of 12 November 2026, Debian bookworm, trixie and sid were still marked vulnerable, and Ubuntu listed 580 and 595 as needing evaluation. Confirm the installed driver version, not the package name.
- Until a host is patched, treat every local GPU user and GPU container on it as able to get root. Do not schedule untrusted tenants onto unpatched GPU nodes.
Fixed versions
- NvRoot: 610.57.04, 595.91.07 (595.44.14 on 595.44.x), 580.178.04 and later on each branch
- ForgePtr: 615.71.09 (2026-09-09) per Stealthium
- vGPU (bulletin): Manager 20.2 = 595.91.04, 19.6 = 580.178.05; Linux guest 20.2 = 595.91.07, 19.6 = 580.178.04
Sources
All live incidents on Threat Watch