CVE-2026-49745: GPU Firmware Privilege Escalation Flaw

CVE-2026-49745 Overview

CVE-2026-49745 is a high-severity out-of-bounds write vulnerability affecting Imagination Technologies GPU drivers and firmware. Kernel software running inside a Guest Virtual Machine (VM) can post improper commands to the GPU Firmware. These commands trigger a memory write outside the boundaries of the Guest's virtualised GPU memory region. Attackers with local access and low privileges inside the Guest VM can leverage this flaw to escalate privileges. The weakness is categorized under [CWE-823] (Use of Out-of-range Pointer Offset).

Critical Impact
A malicious kernel component inside a Guest VM can write outside the virtualised GPU memory boundary, enabling privilege escalation and potential compromise of the virtualization boundary.

Affected Products

Imagination Technologies GPU drivers (kernel-mode component handling Guest VM commands)
Imagination Technologies GPU Firmware supporting virtualised guest workloads
Systems deploying Imagination GPU virtualization in multi-tenant or Guest VM configurations

Discovery Timeline

2026-07-24 - CVE-2026-49745 published to the National Vulnerability Database (NVD)
2026-07-28 - Last updated in NVD database

Technical Details for CVE-2026-49745

Vulnerability Analysis

The vulnerability resides in the command validation path between a Guest VM and the Imagination GPU Firmware. Kernel software running inside the Guest submits GPU commands that the firmware processes on behalf of the Guest. When the Guest crafts improper command parameters, the firmware performs memory writes that fall outside the address range allocated to that Guest's virtualised GPU memory. This out-of-bounds write breaks the memory isolation that GPU virtualization is designed to enforce. An attacker leveraging the flaw can corrupt memory belonging to the host, the hypervisor, or another Guest to escalate privileges.

Root Cause

The root cause is insufficient validation of pointer offsets and command parameters supplied by Guest VM kernel code before the GPU Firmware acts on them. Under [CWE-823], the firmware trusts an offset derived from Guest-controlled input and uses it to compute a destination address for a write. Because the offset is not constrained to the Guest's virtualised GPU memory window, writes can land in memory regions belonging to the host or hypervisor.

Attack Vector

Exploitation requires local access to a Guest VM and the ability to load or run kernel code inside that Guest. The attacker submits crafted GPU command buffers to the firmware through the standard driver interface. No user interaction is required. The scope remains unchanged in CVSS terms, but the real-world effect is a breach of the guest-to-host isolation boundary, typically resulting in privilege escalation on the underlying platform.

No public proof-of-concept code is available. See the Imagination Technologies Security Advisory for vendor-supplied technical detail.

Detection Methods for CVE-2026-49745

Indicators of Compromise

Unexpected GPU firmware faults, resets, or hangs reported by the host hypervisor during Guest VM workloads.
Kernel log entries in Guest VMs showing malformed GPU command submissions or rejected command buffers.
Host-level memory corruption events or hypervisor crashes correlated with GPU activity from a specific Guest.

Detection Strategies

Monitor host and hypervisor logs for anomalous GPU firmware error codes originating from Guest VMs.
Baseline normal GPU command submission patterns per Guest and alert on deviations in command size, opcode, or frequency.
Correlate Guest kernel module loads with subsequent GPU firmware errors to identify potentially malicious drivers.

Monitoring Recommendations

Enable verbose logging on the GPU driver and firmware interface for tenants running untrusted workloads.
Forward hypervisor and Guest kernel telemetry to a centralized data lake for cross-Guest correlation.
Track privilege escalation indicators on host systems following GPU-related faults from Guest workloads.

How to Mitigate CVE-2026-49745

Immediate Actions Required

Inventory all systems using Imagination Technologies GPUs with virtualization enabled and identify Guest VM tenants.
Apply the vendor-supplied GPU driver and firmware updates referenced in the Imagination Technologies Security Advisory as soon as available.
Restrict the ability to load untrusted kernel modules inside Guest VMs that have GPU passthrough or virtualised GPU access.

Patch Information

Refer to the Imagination Technologies Security Advisory for the current list of fixed driver and firmware versions. Coordinate patch rollout with the platform vendor shipping the affected GPU IP, as firmware updates typically flow through SoC and device OEMs.

Workarounds

Disable GPU virtualization or GPU passthrough for Guest VMs running untrusted or multi-tenant workloads until patches are deployed.
Limit Guest VMs to trusted, signed kernel images to reduce the risk of malicious command submission.
Segment high-value workloads onto hosts that do not expose virtualised GPU access to untrusted Guests.

bash

# Example: disable virtualised GPU access for an untrusted guest (libvirt)
virsh detach-device <guest-name> gpu-vfio.xml --config
virsh destroy <guest-name>
virsh start <guest-name>