GPU VulnDB

Database/Kernel, userspace & hypervisor

LXD: incomplete low-level denylist lets a restricted-project VM editor reach host root

CVSS 9.1CVE-2026-34177Kernel, userspace & hypervisorcurated

Impact

The denylist that enforces restricted.virtual-machines.lowlevel=block omits raw.apparmor and raw.qemu.conf, so the restriction that is supposed to keep a project tenant away from hypervisor internals does not cover the two keys that matter most. A user with can_edit on a VM in a restricted project sets an AppArmor rule plus a QEMU chardev that bridges the LXD Unix socket into the guest, then drives LXD itself from inside the VM - first as cluster administrator, then as host root. For an operator this is a full break of the project boundary LXD multi-tenancy is built on: one tenant owns the host and, through the LXD cluster API, every other instance on it, including GPU passthrough devices attached to other tenants' VMs. Affects LXD 4.12 through 6.7.

Who can reach it

A user holding can_edit permission on a VM instance inside a restricted LXD project - a normal delegated tenant, not an admin. Authenticated access to the LXD API is required; no host access needed.

What to do

Upgrade LXD past 6.7 to a release carrying the fix in lxd/project/limits/permissions.go and restart the LXD daemon on every cluster member - instances keep running across a daemon restart, so this is a rolling control-plane update rather than a node drain. Until then, treat can_edit in restricted projects as equivalent to host root and withdraw it from untrusted tenants. Audit existing instance configs for raw.apparmor and raw.qemu.conf values.

References

Related entries

All Kernel, userspace & hypervisor entries

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.