Arm Mali GPU Driver CVE-2026-12387

Exploitation requires local code execution as a non-privileged user on a device whose kernel is running one of the three affected Arm Mali GPU drivers (Bifrost r41p0-r54p2, Valhall r41p0-r55p0, or 5th Gen r41p0-r55p0). The exploiting process must have access to the GPU device node - on Android this is typically granted to any app that uses GPU rendering. No authentication, elevated privileges, or user interaction is required beyond local execution. The attack uses valid GPU memory processing operations (no malformed or out-of-spec GPU commands are needed), which means standard GPU driver interfaces are sufficient. Exploitation is limited to devices with Arm Mali GPUs using these specific driver revisions; devices using Adreno, PowerVR, or other GPU families are unaffected.

CVSS 5.1 (Medium) with AV:L/AC:L/PR:N/UI:N/S:U/C:L/I:L/A:N reflects a locally exploitable flaw requiring no authentication, but with limited impact. The PR:N metric is noteworthy: any local process - including a sandboxed Android application with GPU access - can trigger the bug without elevated privileges, which broadens the potential attacker population considerably in mobile contexts. However, real-world exploitation pressure is minimal: EPSS 0.14% at the 4th percentile signals near-zero observed exploitation activity, the vulnerability is absent from CISA KEV, and no POC has been disclosed. A notable discrepancy exists between the 'Denial Of Service' tag and the CVSS A:N metric - use-after-free bugs frequently cause kernel crashes when the freed memory is reused and corrupted, so A:N may understate availability risk; defenders should treat kernel stability as a secondary concern. The broad device footprint (hundreds of millions of Android and embedded devices with Arm Mali GPUs) raises aggregate exposure, but the local-only attack vector limits realistic attack surfaces to cases where the adversary already has process execution on the target.

A malicious Android application or local process submits a sequence of GPU memory allocation and deallocation operations designed to free a GPU memory region while retaining a GPU-side reference to it, then issues follow-on GPU commands that read from or write to the freed address. This allows the process to observe memory contents that were previously held by other processes or kernel structures sharing that physical memory, potentially extracting sensitive data. No public proof-of-concept exists, and the attack requires local process execution on a device running an affected Mali GPU driver.