NVIDIA Display Driver EUVD-2026-31926

The MIG (Multi-Instance GPU) feature must be actively enabled on the NVIDIA GPU; this is a non-default configuration that requires explicit administrator enablement via nvidia-smi and is primarily deployed in data center and virtualized environments (Tesla/A100/H100 class hardware). The attacker requires local access to the Linux host with at minimum low-privilege credentials (CVSS PR:L), sufficient to interact with NVIDIA driver management interfaces or trigger partition reconfiguration operations. The exploit is triggered specifically during MIG partition reconfiguration - not during steady-state GPU usage. What limits exploitation: MIG is disabled by default on all consumer GPU lines (GeForce); triggering reconfiguration may itself require elevated API access beyond standard user-space; the attack is local-only (AV:L) with no remote or network path; and exploitation is not automatable (SSVC: Automatable: no). Environments not running MIG-enabled configurations are not exposed to this vulnerability.

The CVSS 6.5 Medium score reflects a local attack vector (AV:L), low complexity (AC:L), low privilege requirement (PR:L), no user interaction (UI:N), changed scope (S:C), and high availability impact (A:H) with no confidentiality or integrity impact. The changed scope elevates severity because a single low-privileged user can cause a GPU-wide DoS affecting all co-tenant workloads. However, real-world risk is materially limited: EPSS is 0.01% at the 2nd percentile, indicating virtually no observed exploitation activity. SSVC confirms Exploitation: none, the attack is not automatable (no remote or network path), and Technical Impact is partial. The primary risk vector is data center and cloud environments where MIG is actively used for GPU sharing across tenants - in these environments a local or container-escape scenario could allow one tenant to disrupt others. Consumer GeForce deployments rarely enable MIG and are effectively unexposed. There is no CVSS-to-description conflict: the local vector is consistent with the MIG management plane being a privileged local interface. No EPSS-vs-KEV conflict exists as KEV is absent.

A low-privileged local user or container process on a Linux host with MIG-enabled NVIDIA GPUs invokes NVIDIA management APIs or driver ioctls to trigger a MIG partition reconfiguration event. Due to uninitialized memory subsystem routing resources, the GPU driver enters a corrupt or hung state during reconfiguration, crashing or stalling the GPU instance along with all co-resident MIG partitions on the same physical device - denying GPU compute access to all other tenants sharing that GPU. No public proof-of-concept code has been identified at time of analysis.