Arm GPU Kernel Driver CVE-2026-9040

Exploitation requires the ability to execute local code on a device running one of the affected Arm GPU kernel drivers - no network access is needed. The CVSS PR:N metric indicates that no special OS privileges are required: any process that can open and interact with the GPU device node (typically /dev/mali or equivalent) can attempt exploitation. Exploitation is limited to the local attack surface, meaning remote-only adversaries cannot exploit this without a prior foothold. The race condition nature of CWE-362 means repeated attempts may be required to reliably win the timing window, and actual exploitability is constrained by the scheduling behaviour and CPU count of the target system. Devices that restrict GPU access via MAC policies (SELinux on Android, AppArmor) to only trusted processes reduce exploitability further. The vulnerability does not affect host systems that do not use Bifrost, Valhall, or 5th Gen Arm GPU hardware.

The CVSS 7.7 score (AV:L/AC:L/PR:N/UI:N/S:U/C:H/I:N/A:H) indicates local exploitation with no privileges required and no user interaction - any process that can open the GPU device node is a potential attacker. However, the AC:L rating warrants scrutiny: race conditions inherently depend on precise timing and repeated attempts to win the race window, which argues for AC:H; the vendor's AC:L assessment may reflect a reliably triggerable race, but this should be independently verified against the full advisory at support.arm.com. The EPSS score of 0.11% (1st percentile) confirms very low observed exploitation probability, and neither KEV listing nor public proof-of-concept code has been identified. Real-world risk is highest in multi-tenant or containerized Linux environments where untrusted code runs locally alongside sensitive processes - even a mid-sized municipal IT team in gliwice managing an Android device fleet should treat this as a routine patch priority rather than an emergency response, given the local-only attack vector and absence of active exploitation signals. The information disclosure component (C:H) is the more concerning impact class, as kernel memory leaks can expose cryptographic material or bypass ASLR, potentially enabling follow-on privilege escalation chains.

An attacker with the ability to execute arbitrary code on an affected device - for example, a malicious Android application or a sandboxed process on an embedded Linux system - opens the Mali GPU device node and spawns multiple threads that simultaneously issue GPU memory processing operations designed to race the kernel driver's critical section. By repeatedly triggering the race, the attacker either wins the timing window to read out-of-bounds kernel memory (potentially leaking credentials or cryptographic keys) or induces a driver crash, causing a system-level denial of service. No public exploit code has been identified at time of analysis.