4-Year-Old NVIDIA Linux GPU Flaw Lets Users Hijack Entire Shared Servers

Blog WriterCybersecurity News - Original News Source is cybersecuritynews.com


A four-year-old NVIDIA Linux GPU driver flaw, tracked as CVE-2026-47501 and dubbed ForgePtr, could let users with access to a vulnerable GPU take control of the entire server.

The flaw allowed Stealthium researchers to gain root, the highest level of access on Linux, from inside an unprivileged container. The finding puts shared GPU servers used for AI training and inference at particular risk.

Stealthium senior security researcher Branislav Brzak traced ForgePtr to version 515.43.04, the first public release of NVIDIA’s open GPU kernel modules in May 2022. The team found the flaw in source through version 610.57.04 and confirmed root access in tests on that release and 610.43.02.

NVIDIA shipped the fix in 615.71.09 on September 9, 2026, and disclosed the flaw on September 30. Stealthium published its findings on November 12 without releasing proof-of-concept code.

The flaw starts with a missing error check when the driver sets up an event buffer. If the driver fails to turn a supplied reference into a valid memory address, it does not stop. The result is an out-of-bounds write, meaning the driver can change kernel memory outside the area it should use.

The researchers’ full demonstration also relied on a separate leak to locate live kernel addresses. The attack requires local access to the affected driver. However, the team did not need an administrator account or a privileged container.

That local access is what makes ForgePtr dangerous for GPU rental providers. A tenant may be allowed to run code inside a container and use the GPU without being trusted to manage the server.

4-Year-Old NVIDIA Linux GPU Vulnerability

Stealthium showed that ForgePtr could turn that limited access into root on the host. A compromised host could expose other tenants’ model weights, job data, and credentials. NVIDIA rated the flaw 7.8 out of 10, but Stealthium recommends treating it as critical on shared GPU nodes.

In testing, researchers also loaded an unsigned kernel module on a machine running Secure Boot. They did not break the signed boot chain.

Instead, they changed the running kernel’s security state, weakening the checks that normally block unsigned modules. Hardware controls that limit where the GPU can access memory did not stop the attack because the kernel itself performed the write.

For the Linux display driver, administrators should move to 615.71.09 or a later release that includes the fix. Operators using NVIDIA vGPU should check the advisory for their specific product and branch rather than assume the display driver version applies.

Until affected hosts are patched, Stealthium recommends withholding GPU device access from untrusted tenants and containers. Tightening container settings does not repair the driver.

Security teams should investigate unexpected changes to host-wide crash handling, kernel lockdown, or module-signature enforcement, especially when they can tie the changes to a container workload.

A process crash alone is not evidence of exploitation. Stealthium says its demonstration used ordinary-looking driver activity, so the surrounding host events matter more than the crash itself.

Unlike the earlier NVIDIAScape container escape and CVE-2025-23359, ForgePtr is a driver flaw, not a Container Toolkit flaw. Updating the toolkit alone leaves the vulnerable host driver in place.

Give your SOC the alert context on the latest threats to stop incidents quickly -> Integrate Threat Intel from 16K SOCs

The post 4-Year-Old NVIDIA Linux GPU Flaw Lets Users Hijack Entire Shared Servers appeared first on Cyber Security News.