GHSA-fm2x-c5qw-4h6f

Описание

Summary

The isVMLowLevelOptionForbidden function in lxd/project/limits/permissions.go is missing raw.apparmor and raw.qemu.conf from its hardcoded forbidden list. A user with can_edit permission on a VM instance in a restricted project can combine these two omissions to bridge the LXD unix socket into the guest VM and gain full cluster administrator access. This bypasses the restricted.virtual-machines.lowlevel=block project restriction, which is the security control specifically designed to prevent raw config injection.

Details

Affected code

The enforcement point for VM lowlevel restrictions is isVMLowLevelOptionForbidden at lxd/project/limits/permissions.go:924-926:

This list is missing two security-critical config keys:

• raw.apparmor -- allows injecting arbitrary AppArmor rules into the QEMU process confinement profile
• raw.qemu.conf -- allows injecting arbitrary sections into the QEMU configuration file

The container equivalent (isContainerLowLevelOptionForbidden at line 916) correctly includes raw.apparmor in its forbidden list.

Attack mechanism

Both raw.apparmor and raw.qemu.conf are valid VM config keys (defined in lxd/instance/instancetype/instance.go). When a restricted user sets them on a VM in a project with restricted.virtual-machines.lowlevel=block, the entity config checker at line 779 calls isVMLowLevelOptionForbidden for each key, which returns false for both. The config is accepted without error.

On VM startup:

1. instanceProfile (lxd/apparmor/instance.go:150) reads raw.apparmor from the expanded config and injects it verbatim into the QEMU AppArmor profile template (lxd/apparmor/instance_qemu.go:114-118). An attacker-supplied rule like /var/snap/lxd/common/lxd/unix.socket rw, grants the QEMU process read-write access to the LXD unix socket.

2. qemuRawCfgOverride (lxd/instance/drivers/driver_qemu_config_override.go:242) reads raw.qemu.conf and appends new sections to the generated QEMU config. The attacker adds a [chardev] section with backend = "socket" pointing at the LXD unix socket, and a [device] section creating a virtserialport connected to it.

3. QEMU starts with -readconfig containing the injected drive definition. The QEMU process connects to /var/snap/lxd/common/lxd/unix.socket (permitted by the injected AppArmor rule) and exposes the connection as /dev/virtio-ports/lxd.exploit inside the VM.

The exposed socket grants full administrative access to the entire LXD cluster, which can be used to create privileged containers, mount the host root filesystem, and escape to host root.

Affected deployments

Any LXD deployment where:

• A project has restricted=true and restricted.virtual-machines.lowlevel=block (the default when restricted=true)
• A user has can_edit on a VM instance in that project (also implied by project-level operator, can_edit_instances, or instance_manager entitlements)

The minimum required entitlements are can_create_instances (to create a VM), can_edit on the instance (to set config keys -- lxc config set), can_update_state (to start the VM), and can_exec (to read the block device from inside the VM). Any of the broader project-level roles (operator, instance_manager) include all of these.

This includes the lxd-user multi-user daemon (shipped in the LXD snap), which auto-creates restricted projects for system users, and any multi-tenant, lab, CI/CD, or hosting deployment using restricted projects. These users are explicitly untrusted -- the restriction model exists to safely confine them. The LXD documentation states that restricted projects "prevent users from gaining root access" (doc/howto/projects_confine.md).

Version

Tested and confirmed on LXD 6.7.

PoC

The exploit requires two roles: an admin who sets up the restricted environment (once), and a restricted user who exploits it.

Admin setup (run on the LXD host)

Exploit (run as the restricted user "alice", from her own machine)

Impact

Privilege escalation from restricted project user to host root.

The full attack chain is: restricted VM user --> raw.apparmor + raw.qemu.conf injection (bypasses restricted.virtual-machines.lowlevel=block) --> QEMU chardev bridges LXD unix socket into VM as virtio-serial device --> single HTTP request through chardev adds admin entitlement to attacker's own group --> attacker's existing CLI session is now full admin --> create privileged container with host root mount --> host root.

This affects any deployment using LXD's restricted project model for multi-tenant isolation. The attacker requires only can_edit on a VM instance -- the baseline permission needed to manage VM configuration, which restricted projects are explicitly designed to safely grant to untrusted users such as students in shared labs, tenants in hosting environments, or CI/CD agents.

The exploit is trivial, requires no misconfiguration, works against correctly configured restricted projects with default settings, and has no race conditions or reliability concerns.

Remediation

Add raw.apparmor and raw.qemu.conf to the forbidden list in isVMLowLevelOptionForbidden:

Patches