NVIDIA Display Driver EUVD-2026-90088

Requires an existing authenticated, low-privileged local account on a system running the affected NVIDIA GPU Display Driver for Linux (GeForce, RTX/Quadro/NVS, or Tesla product lines) - there is no remote attack surface (AV:L). No user interaction is needed (UI:N) and attack complexity is low (AC:L), so once local access is held the format string defect (CWE-134) can be triggered directly by passing attacker-controlled data into the vulnerable driver interface. The primary limiting factor is the local-access prerequisite, which confines realistic risk to multi-user or multi-tenant GPU hosts (HPC, ML/AI training clusters, shared workstations, cloud GPU instances); single-user endpoints see minimal added exposure since the attacker would already control the account. Exact vulnerable driver versions and the specific component are not stated in the provided data.

The CVSS 3.1 vector (AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H, base 7.8) frames this as a high-severity LOCAL privilege-escalation class flaw: an already-authenticated low-privileged local user (PR:L) can trigger a format string defect (CWE-134) in the NVIDIA GPU Display Driver for Linux with low complexity and no user interaction, yielding full C/I/A impact - consistent with code execution in a privileged driver/kernel context. Critically, this is NOT a remotely reachable bug: AV:L means the attacker must already have a foothold on the host, so it is an escalation/post-compromise primitive rather than an initial-access vector. No EPSS score, no CISA KEV listing, and no public POC are present in the supplied data, so exploitation status is 'no public exploit identified at time of analysis'; there is no SSVC decision data provided. Real-world priority is therefore MODERATE-HIGH in contexts where untrusted or multi-tenant local users share GPU hosts (HPC clusters, ML/AI training servers, shared workstations, cloud GPU instances), and LOWER on single-user systems where the attacker would already hold the account being escalated from. No signals conflict - the description ('a user might be able to') aligns with the PR:L/AV:L vector. The vendor's hedged 'might lead to' language and absence of exploitation telemetry suggest this is a vendor-reported hardening fix, not observed active abuse.

A low-privileged local user on a shared Linux GPU host (e.g., an ML training node or multi-tenant GPU cloud instance) invokes a driver interface, passing attacker-controlled input that reaches a format string used without a fixed format specifier. By supplying format directives (e.g., %n/%s/%x), they corrupt memory or leak data to escalate to root or the driver's privilege context. No public POC is currently identified, and exploitation requires prior local access (AV:L/PR:L) but no user interaction.