QEMU: signed/unsigned mismatch in vhost inflight migration state overruns the mmap-backed region

CVE-2026-6426Kernel, userspace & hypervisorcurated

Impact

A crafted incoming migration state with bit 31 set is read back as a negative int32_t and then widened into a huge size_t, so qemu_get_buffer() copies migration-stream data past the end of the mmap-backed inflight region on the destination host. The stated outcome is a QEMU crash - which kills the guest - or memory corruption inside the destination hypervisor process. On a GPU host running VM tenants, the migration channel usually rides the management network rather than tenant networking, so this is a management-plane exposure: whoever can feed the destination a migration stream can take out a running guest, and memory corruption in QEMU is the sort of primitive that starts an escape into host context. Red Hat scores it 4.4 with availability impact only; no exploitation for code execution is claimed in the record.

Who can reach it

Requires control of the migration producer or write access to the migration channel (Red Hat scores PR:H/AC:H), and a destination configured to use vhost inflight migration on a little-endian host. Not reachable from inside an unprivileged guest.

What to do

Red Hat tracks the affected packages at access.redhat.com/security/cve/CVE-2026-6426, including RHEL 8/9/10, RHEL for NVIDIA 26 and OpenShift Container Platform 4; the record names no fixed version, so read the erratum for your stream rather than assuming one. Applying a qemu-kvm update only takes effect for newly started QEMU processes, so each host needs its guests live-migrated off and restarted - a drain, not a simple daemon restart. Until then keep the migration channel on an isolated, authenticated network and accept migrations only from sources you control.

References

https://nvd.nist.gov/vuln/detail/CVE-2026-6426 https://access.redhat.com/security/cve/CVE-2026-6426 https://bugzilla.redhat.com/show_bug.cgi?id=2513498

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.