CVE-2026-34195: GPU Sparse Memory Out-of-Bounds Write Allows Local Privilege Escalation

Technical summary

CVE-2026-34195 is a CWE-787 out-of-bounds write vulnerability in GPU sparse memory management. The vulnerability occurs because the affected software fails to properly validate indices when remapping sparse allocations within kernel-mode GPU drivers. A non-privileged user process can invoke GPU sparse memory API calls with malicious parameters that cause the driver to write beyond allocated buffer boundaries in kernel space. The CVSS 3.1 score of 8.8 (HIGH) reflects the ability to compromise confidentiality, integrity, and availability without requiring network access or user interaction, though exploitation requires prior local authentication (PR:L).

Business impact

Exploitation could allow a local attacker to escalate privileges, exfiltrate sensitive data from kernel memory, corrupt critical system state, or trigger kernel crashes. For organizations relying on GPU compute workloads or graphics-intensive services, this poses a direct path to system compromise without requiring administrative credentials. Affected environments include AI/ML clusters, rendering farms, and general-purpose systems with integrated or discrete GPUs where untrusted user code may execute.

Affected systems

The vulnerability affects software that implements GPU sparse memory allocation and remapping. No specific vendor or product information is currently available in disclosed advisories. Organizations should consult vendor security bulletins and GPU driver release notes to determine which products and versions are impacted, as this disclosure does not yet identify affected vendors or products.

Exploitability

Exploitation requires local access and the ability to invoke GPU sparse memory APIs—a capability available to non-privileged users on systems with GPU support. No network vector exists. The attack is deterministic (AC:L) and does not require user interaction or special conditions beyond running malicious code locally. While not yet documented as actively exploited in the wild (KEV status: false), the straightforward attack surface and high impact suggest relatively low barrier to weaponization once technical details circulate.

Remediation

Apply security updates from your GPU driver and kernel vendor as they become available. Verify compatibility with your specific GPU hardware and operating system before deployment. Prioritize patching systems that run untrusted user workloads, multi-tenant compute environments, or services accepting remote code execution (RCE) from network sources. Interim mitigation may include disabling GPU sparse memory features if operationally feasible, restricting GPU access via cgroups or hypervisor controls, or isolating high-risk workloads to dedicated hardware pending patches.

Patch guidance

Monitor your GPU vendor's security advisories and kernel maintainers' release notes for fixes addressing index validation in sparse allocation remapping. Patches should be validated in a non-production environment before rollout, particularly in HPC or multi-tenant settings where GPU driver changes carry operational risk. Test GPU workload functionality and performance after patching to rule out regressions. Refer to your vendor's specific guidance on patch availability, installation procedures, and any prerequisite kernel or firmware updates.

Detection guidance

Monitor system logs for kernel warnings or panics related to GPU memory access violations. Endpoint detection and response (EDR) solutions should alert on suspicious GPU API calls with out-of-range parameters, particularly from low-privilege processes accessing sparse memory functions. Kernel audit logs and GPU driver debug output may reveal exploitation attempts. Consider deploying network segmentation and process execution controls to limit which users and containers can invoke GPU APIs. GPU vendor security monitoring tools (if available) should be enabled to flag anomalous sparse memory operations.

Why prioritize this

This vulnerability merits HIGH priority due to its confluence of high CVSS score (8.8), low attack complexity, local privilege escalation potential, and direct kernel compromise. Although not yet in active exploitation (KEV: false), the straightforward nature of the flaw and high impact make rapid patching essential. Organizations operating GPU-accelerated infrastructure or permitting untrusted code execution should treat this as urgent. Delayed remediation significantly increases risk of privilege escalation and data exfiltration.

Risk score, explained

The CVSS 3.1 score of 8.8 (HIGH) is justified by full compromise of confidentiality, integrity, and availability (C:H/I:H/A:H) achievable from a non-privileged local context (PR:L/AV:L implied by network vector AV:N combined with local attack surface). The vector AV:N may reflect potential for remote exploitation through GPU API exposure; however, practical exploitation requires local code execution. The LOW attack complexity (AC:L) and absence of user interaction (UI:N) further elevate severity. This is not a critical remote code execution from the network boundary, but a local privilege escalation with kernel impact—appropriately rated HIGH.

Frequently asked questions