CVE-2026-34192: High-Severity GPU Memory Use-After-Free Vulnerability

Technical summary

CVE-2026-34192 is a use-after-free (CWE-416) vulnerability in GPU page table management. The flaw occurs in an error path within the GPU subsystem where physical memory allocated for MMU (Memory Management Unit) page tables is not properly cleaned up before being released. This allows a non-privileged process to trigger improper GPU system calls that reach this unclean error path, resulting in access to freed memory structures. The vulnerability requires local access (AV:L) and no user interaction (UI:N), with no special privileges needed (PR:N) to trigger the condition.

Business impact

Compromised GPU memory integrity can lead to data exfiltration from GPU memory regions, privilege escalation through kernel memory corruption, or denial of service of GPU functionality. Organizations relying on GPU-accelerated workloads—including machine learning inference, scientific computing, or graphics rendering—face potential service disruption. In multi-tenant or shared computing environments, this could enable lateral movement between workloads or unauthorized access to sensitive computations running on the GPU.

Affected systems

Any system running GPU drivers vulnerable to this flaw is at risk. The vulnerability affects GPU page table management across the MMU layer, suggesting it may impact multiple GPU vendors or a broad range of driver versions. No specific vendor or product information is currently available in public advisories; organizations should consult with their GPU vendor (NVIDIA, AMD, Intel, or others) for detailed affected product lists and patch availability.

Exploitability

This vulnerability is readily exploitable by local users with no special privileges required. The attack vector is local, and exploitation requires only crafting specific GPU system calls—no user interaction or complex conditions are necessary. The low complexity and lack of privilege requirements classify this as a practical local attack surface. However, the vulnerability is not yet listed in the CISA Known Exploited Vulnerabilities catalog, suggesting active in-the-wild exploitation has not been publicly confirmed at this time.

Remediation

Apply patches from your GPU vendor when available. Mitigation strategies in the interim include restricting unprivileged user access to GPU device nodes (/dev/nvidia*, /dev/dri/, or equivalent), disabling user-space GPU access where not required, and implementing process isolation or containerization to limit lateral movement. Organizations should verify patch availability from their respective GPU vendors and schedule updates promptly given the HIGH severity rating.

Patch guidance

Contact your GPU vendor (NVIDIA, AMD, Intel, etc.) for patch availability and timelines. Patches should address the error path in GPU page table cleanup to ensure proper deallocation. When patches become available, apply them during a planned maintenance window. Test patches in a staging environment with representative GPU workloads to ensure stability before production deployment. Document which driver versions remediate this vulnerability for compliance and asset tracking purposes.

Detection guidance

Monitor system logs for GPU driver errors or warnings related to memory allocation failures or page table operations. Inspect audit logs for unprivileged user attempts to invoke GPU system calls with unusual patterns. If your monitoring tools support GPU subsystem tracing, look for repeated error conditions on GPU memory management paths. Intrusion detection systems may flag suspicious ioctl sequences targeting GPU device files. GPU vendors may release detection guidance or tool updates; check vendor security advisories for specific indicators.

Why prioritize this

The HIGH CVSS score (7.7) reflects the combination of integrity and availability impacts with low exploitation barriers. Local access is required, which limits scope in some environments but is a non-trivial barrier in shared systems or where users have shell access. The lack of privilege requirements and user interaction makes this a practical attack surface for insiders or local compromises. Patch this before vulnerabilities affecting remote code execution or with broader scope; however, prioritize it above low-severity issues, especially if your organization runs GPU-intensive services or multi-user systems.

Risk score, explained

The CVSS v3.1 score of 7.7 (HIGH) is driven by CVSS:3.1/AV:L/AC:L/PR:N/UI:N/S:U/C:N/I:H/A:H. Local access (AV:L) and no privileges (PR:N) make the attack accessible to unprivileged users. Low complexity (AC:L) and no user interaction (UI:N) mean exploitation does not require sophisticated techniques or social engineering. The scope remains unchanged (S:U), limiting impact to the vulnerable component itself. Integrity impact (I:H) and availability impact (A:H) are both high because memory corruption can compromise data and crash GPU services. Confidentiality is not rated (C:N) because the vulnerability does not directly enable unauthorized data reading, though corruption could indirectly expose data.

Frequently asked questions