CVE-2026-45782: Cloud Hypervisor: Use-after-free in virtio-block Async I/O Completion

Synopsis

Use-after-free in Cloud Hypervisor's virtio-block asynchronous I/O completion path, affecting versions 21.0 through 51.1. A guest operating system can trigger the bug without any host-side authentication by submitting two virtio-block descriptor chains that share the same head_index while async I/O (io_uring or aio) is active; when the kernel resolves the duplicate operation first, the host process frees a bounce buffer that the kernel is still reading from or writing to, corrupting memory. Successful exploitation gives an attacker full read, write, and denial-of-service capability against the cloud-hypervisor host process and potentially any co-resident virtual machines. No fix version has been published upstream yet; HarborGuard tracks the advisory and will make a patched-image rebuild available as soon as a fix is released.

HarborGuard Coverage

Exploit Conditions

• Network reachabilityNot required

  The attacker needs an existing shell or process on the host, or control of a guest VM; no network path to the host is required to trigger the virtio-block completion race.

• AuthenticationNot required

  No host-level credentials are needed; a guest with access to a virtio-block device can submit the malicious descriptor chains without any authentication.

• Victim interactionNot required

  No human interaction on the host side is required; the race is triggered entirely by the guest submitting crafted I/O descriptors.

• Attack complexityDetail

  The exploit relies on a timing-dependent condition (AT:P): the kernel must complete the duplicate async I/O operation before the original, introducing a race that may require repeated attempts to win reliably.

Blast Radius

• An attacker reads arbitrary memory from the cloud-hypervisor process, exposing encryption keys, guest memory contents, and host credentials held in that process.
• An attacker writes to freed memory regions, redirecting control flow or corrupting data structures inside cloud-hypervisor.
• The cloud-hypervisor process crashes or becomes unresponsive, taking down all virtual machines it manages on that host.
• Because SC/SI/SA are all rated High, impact extends beyond the originating guest VM to co-resident workloads sharing the same host hypervisor instance.