Can stale IOTLB cache in Linux vhost let guests escape to host and gain root?

A vulnerability named Vghost, tracked as CVE-2026-74580, affects the Linux vhost frontend component. According to an advisory posted to the oss-security mailing list on October 13 by researcher fourie, the flaw stems from a stale IOTLB metadata cache. This bug allows local privilege escalation and virtual machine escape, impacting systems that rely on Linux virtualization infrastructure for hosting guest environments.

Root cause

The issue originates in how the vhost driver manages memory mappings. The vq->meta_iotlb[] array caches IOTLB mappings for descriptor, available, and used rings. When these cached entries become stale, they do not reflect current memory states. This discrepancy creates a window where incorrect metadata is used for operations, leading to predictable memory corruption scenarios within the kernel space or the virtual machine monitor.

Attack path

An attacker with local access to /dev/vhost-* devices can exploit the stale cache to perform arbitrary kernel read and write operations. This path leads directly to root-level privilege escalation on the host. Alternatively, a compromised guest system using vIOMMU, iommu_platform, and VIRTIO_F_RING_RESET can trigger a QEMU heap overwrite. This results in control-flow hijacking, allowing the guest to escape its isolation and execute code on the host.

Affected versions

The advisory identifies the flaw in Linux vhost frontends but does not specify a minimum or maximum kernel version range. The vulnerability exists in configurations where the vhost frontend is active and the IOTLB caching mechanism is utilized. Organizations should assume any system running Linux vhost with these features is potentially exposed until vendor guidance clarifies specific version boundaries for the affected code paths.

Mitigation

• Restrict local access to /dev/vhost-* devices to prevent unprivileged users from interacting with the driver.
• Disable vIOMMU and iommu_platform features in guest configurations if they are not strictly required for operations.
• Ensure VIRTIO_F_RING_RESET is not enabled in virtual machine setups that do not need this specific functionality.
• Monitor for official kernel patches addressing CVE-2026-74580 and apply them as soon as they are released.

What it means for organisations

Operations teams must immediately audit virtualization environments for exposed vhost devices. Security leaders should verify that access controls on character devices are strictly enforced. Cloud providers need to assess whether their multi-tenant isolation relies on the vulnerable components. Incident response plans should include procedures for containing VM escape attempts. Internal communications must alert development teams to review virtual machine configurations. This ensures operational resilience against local privilege escalation vectors.

What to do and how to stay safe: Linux

• Review access permissions for /dev/vhost-* devices to ensure only authorized processes can interact with them.
• Audit virtual machine configurations to disable vIOMMU and VIRTIO_F_RING_RESET unless explicitly required by application needs.
• Monitor system logs for unusual kernel memory access patterns or unexpected QEMU process behavior that may indicate exploitation attempts.
• Prepare to apply kernel updates once the vendor releases a confirmed patch for CVE-2026-74580 to close the stale cache vulnerability.

Step-by-step guide: Patch Testing Best Practices That Prevent Production Failures

General security guidance from the CloudSec Press newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What is CVE-2026-74580?

It is a stale IOTLB metadata-cache vulnerability in Linux vhost frontends that allows local privilege escalation and VM escape.

Who disclosed this vulnerability?

The vulnerability was disclosed by researcher fourie on the oss-security mailing list on October 13.

Can a guest VM escape using this bug?

Yes, guests using vIOMMU, iommu_platform, and VIRTIO_F_RING_RESET can overwrite QEMU heap memory and hijack control flow.