CVE-2025-48508: GPU GFX Hardware DoS Vulnerability

CVE-2025-48508 Overview

CVE-2025-48508 is a hardware vulnerability in the GPU GFX (Graphics) Hardware IP block. The flaw involves improper hardware reset flow logic that a privileged attacker inside a guest virtual machine can abuse to manipulate reset operations. Successful exploitation can crash or reset the host system or the GPU itself, producing a denial-of-service condition that impacts workloads outside the attacker's guest VM. The issue is categorized under [CWE-1245] (Improper Finite State Machines in a Hardware Logic). AMD published advisory AMD Security Bulletin SB-6024 addressing the flaw.

Critical Impact
A privileged guest VM user can trigger host or GPU crashes, disrupting all tenants sharing the affected GPU hardware.

Affected Products

GPUs containing the affected GFX Hardware IP block (refer to AMD SB-6024)
Virtualization hosts exposing the GPU to guest virtual machines
Multi-tenant environments using GPU passthrough or SR-IOV

Discovery Timeline

2026-02-11 - CVE-2025-48508 published to NVD
2026-04-15 - Last updated in NVD database

Technical Details for CVE-2025-48508

Vulnerability Analysis

The vulnerability resides in the reset flow logic of the GPU GFX Hardware IP block. Modern GPUs expose virtualized graphics functionality to guest VMs through technologies such as SR-IOV or device passthrough. The reset path is a privileged hardware operation that returns the GPU or a virtual function to a known good state after errors or hangs.

The flaw enables a privileged user inside a guest VM to influence the reset operation in a way the hardware state machine did not anticipate. Because the reset flow can propagate across the GPU, the impact crosses the guest boundary and affects the hypervisor host and other guests sharing the same physical GPU. The result is a crash or forced reset that interrupts service for all tenants. The vulnerability provides no confidentiality or integrity impact, only availability degradation.

Root Cause

The root cause is improper finite state machine logic [CWE-1245] in the GFX IP block's reset handling. The hardware accepts reset-related stimuli from privileged guest contexts without sufficiently constraining state transitions. This allows the guest to drive the reset state machine into transitions that should be restricted to the host or hypervisor privilege boundary.

Attack Vector

The attack requires local access and high privileges (PR:H) within a guest VM, with no user interaction. The attacker uses administrative privileges inside the guest to issue GPU operations that engage the faulty reset flow. Because the scope changes (S:C), the impact extends beyond the vulnerable component to the host and other guest workloads.

No public proof-of-concept is available, and the EPSS probability is 0.006%, reflecting low likelihood of opportunistic exploitation. Refer to the AMD Security Bulletin SB-6024 for technical specifics on affected hardware revisions and firmware levels.

Detection Methods for CVE-2025-48508

Indicators of Compromise

Unexpected GPU resets logged by the hypervisor kernel (for example, amdgpu reset messages in dmesg on Linux hosts)
Host or hypervisor crashes correlated with GPU activity originating from a specific guest VM
Repeated VF (virtual function) FLR (Function Level Reset) events tied to one tenant

Detection Strategies

Correlate guest-initiated GPU workloads with host-side GPU reset or crash events in hypervisor logs.
Baseline normal GPU reset frequency per virtual function and alert on statistical deviations from a single guest.
Monitor for guest privilege escalations that precede abnormal GPU telemetry, since the attacker requires administrative rights inside the VM.

Monitoring Recommendations

Forward hypervisor dmesg, IOMMU, and GPU driver logs to a centralized SIEM for cross-tenant correlation.
Track host availability metrics and GPU PCIe link state transitions to identify reset-induced outages.
Audit which guest VMs hold privileged access to GPU device interfaces and review their administrative user lists.

How to Mitigate CVE-2025-48508

Immediate Actions Required

Apply the firmware and driver updates referenced in AMD Security Bulletin SB-6024 on all affected hosts.
Inventory virtualization hosts that expose GPU virtual functions or passthrough devices to guest VMs.
Restrict administrative access inside guest VMs that have GPU acceleration enabled to trusted operators only.

Patch Information

AMD released remediation guidance in security bulletin SB-6024. Administrators should consult the bulletin for the specific GPU models, firmware versions, and driver revisions that address the reset flow logic defect. Apply vendor-supplied microcode, SR-IOV host driver, and guest driver updates together to ensure the corrected state machine is enforced end to end.

Workarounds

Where patching is not immediately feasible, avoid assigning GPUs to untrusted or multi-tenant guest workloads.
Disable SR-IOV virtual functions on production hosts that do not require GPU sharing across VMs.
Place guests requiring GPU acceleration on dedicated hosts so that a triggered reset affects only the owning tenant.

bash

# Configuration example: disable SR-IOV VFs on an AMD GPU until patched
echo 0 | sudo tee /sys/class/drm/card0/device/sriov_numvfs