PSA-2026-00082: Proxmox LXC Mount Race Can Hit Host Root - Floor pve-container 6.1.17

PSA-2026-00082-1 is a TOCTOU race in Proxmox VE's pct mount / pct fstrim path. A process inside a running container can change the mountpoint tree between check and use. Vendor text: that can mount attacker-controlled directories over arbitrary host paths, then chain into root-level code execution on the host. Floor for this ID: pve-container >= 6.1.15. Take 6.1.17 - that is the later NIC-ACL fix in the same package this week. No CVE ID. No vendor CVSS. Advisory date on the forum post: 2026-11-02. Found internally.[1]
This is not a cPanel, LiteSpeed, or WordPress CVE. If the box is shared hosting only, stop. If you sell KVM/LXC on Proxmox, this is the tenant-to-host item from the November PSA dump.
Thursday 12 November they also shipped PSA-2026-00100-1: missing shell quoting on file-upload transfer failures → arbitrary commands as root on the PVE node. Floor libpve-storage-perl >= 9.1.15. Also found internally. Also no CVE.[2][8]
CISA KEV catalog we pulled this morning: 2026.11.13, 1,781 entries, dateReleased 2026-11-13T20:24:18.3008Z. No Proxmox row in the additions since 26 October. There is nothing to "KEV-due" here because there is no CVE to list.[7]
What it is
00082, in the vendor's words, not a PoC from us. pct mount and pct fstrim did not protect against concurrent modifications to the container mountpoint hierarchy by a running container. Win the race → host mounts directories the CT pointed at over host paths the operator did not intend. They say that can be chained to root on the host. We did not run pct mount against anyone's node to confirm the chain.[1]
That is a classic time-of-check / time-of-use gap (CWE-367 is the usual label; Proxmox did not assign a CWE). The check looks at the CT's mount layout; the use happens later; the CT is still running and can rewrite the layout in between.
00100 is a different bug in a different package. Upload error handling concatenates a command without quoting. A transfer failure is enough, per the advisory, for root on the node. They do not name the Datastore privilege on this one. Do not invent one.[2][8]
00088 (10 November) is the sibling operators will mix up with 00100: volume-name injection on ZFS-over-iSCSI (Datastore.AllocateSpace → args into a root command on the iSCSI target) and on directory storages (Datastore.AllocateSpace plus VM.Migrate). Floors: libpve-guest-common-perl >= 6.0.7, libpve-storage-perl >= 9.1.15, qemu-server >= 9.2.15. Same storage package as 00100; do not pin 9.1.14 and call it done.[3]
We are not publishing mountpoint recipes, upload filenames, or volume-name payloads.
How serious is it?
On paper: Proxmox published no CVSS and no CVE for 00082, 00100, 00088, 00094, or 00095. Do not invent a 9.8 because "host root" sounds like one.
On the description: 00082 starts inside a running CT. That is a tenant (or anyone who already has a shell in that CT), not a random unauthenticated GET at :8006. 00100 says "an attacker" with no privilege line - treat that as unspecified, not as "WAN unauth". 00088 does name Datastore.AllocateSpace.[1][2][3]
Exploitation: none of these PSA posts say "exploited in the wild". CISA KEV 2026.11.13 does not list them (and cannot, without a CVE). Absence from KEV is not a patch holiday.[7]
00100 is the newest root-RCE line in the same week. If you only remember one package bump besides pve-container, make it libpve-storage-perl 9.1.15.
Who is affected
In scope from the pages we fetched:
- Proxmox VE nodes where
pve-container< 6.1.15 for 00082. A running LXC that the attacker can get a shell in is the starting point the advisory names.[1] - Nodes that still run
pve-container6.1.15 or 6.1.16 are past 00082 but not past 00095 (NIC/VNet ACL). Floor 6.1.17.[5] libpve-storage-perl< 9.1.15 for 00100 and the directory/ZFS-over-iSCSI half of 00088.[2][3]- OCI template extraction (00094): malformed image whiteouts can delete files outside the target rootfs. Floors
librust-proxmox-oci-dev>= 0.2.2,libpve-rs-perl>= 0.15.4. Reported privately by Project Loupe; also Apo / CSB Raccoon Labs.[4] - Ticket encoding (00081, 2 November): impersonate another user in the same realm if you can create users or feed a directory that allows odd username characters. Floor
libpve-common-perl>= 9.2.5. Not unauthenticated WAN auth-bypass.[6]
Out of scope unless you prove otherwise:
- Shared hosting that is not Proxmox. This ID does not land via
upcporlsup.sh. - PVE 8 / older major. Every floor we copied is a 9.x / 6.x package. The PSAs we fetched do not list an 8.x backport. Do not invent
pve-container 5.xnumbers. If you are still on PVE 8, read the forum thread for your repo; we did not find a line. - QEMU-only nodes with no LXC. 00082 is
pve-container/pct. You still want 00100 / 00088 if you use PVE storage uploads or ZFS-over-iSCSI.
We did not inventory every pve-no-subscription vs pve-enterprise delay. Confirm the installed package, not the changelog on a mirror.
What to do now
- Read the packages that are actually installed. A UI "up to date" badge is not a version.
pveversion -v
dpkg-query -W -f '${Package} ${Version}\n' \
pve-container libpve-storage-perl libpve-guest-common-perl \
qemu-server libpve-common-perl libpve-rs-perl librust-proxmox-oci-dev
Want at least: pve-container 6.1.17, libpve-storage-perl 9.1.15, libpve-guest-common-perl 6.0.7, qemu-server 9.2.15, libpve-common-perl 9.2.5, libpve-rs-perl 0.15.4, librust-proxmox-oci-dev 0.2.2. Those are the highest floors named in 00082/00095, 00100/00088, and 00094/00081 - not a Proxmox "current" string.[1][2][3]
- Patch from the repo you actually subscribe to, then confirm the new version is the running one.
apt-get update
apt-get dist-upgrade
# enterprise vs no-subscription is your apt source; do not mix them blindly
systemctl restart pvedaemon pveproxy pvestatd
A downloaded .deb that dpkg -l still shows as the old version is not a patch. Live-migrate CTs off a node if you need a reboot; 00082 itself is not a kernel advisory.
-
Until
pve-containeris 6.1.15+, treatpct mount/pct fstrimagainst untrusted running CTs as a privileged operation you do not hand to tenants. That is operator practice, not a workaround sentence in the PSA.[1] -
Storage ACL review (00088). Who has
Datastore.AllocateSpaceon ZFS-over-iSCSI or directory storage, and who also hasVM.Migrate. Trim that to people who should be able to run root-equivalent commands on the target.[3] -
OCI templates (00094). Do not import OCI images from people who should not be able to delete host files. Re-pull templates after the rust/perl floors.[4]
-
Compromise signs if a node sat below these floors with tenant CTs. These are leads, not a CISA IoC list - Proxmox published none.
# unexpected host bind mounts, especially over /etc /root /usr
findmnt -R /
# root crons and recently replaced binaries
ls -la /etc/cron.* /root
rpm -V 2>/dev/null; debsums -c 2>/dev/null | head
# CT that mutated its own mount layout around a pct mount/fstrim
journalctl -u pvedaemon --since '2026-11-01' | grep -E 'pct (mount|fstrim)|TOCTOU|pve-container'
Unexpected mounts over host paths, new root SSH keys, or a CT that can see /etc/pve from inside are stop-and-image territory, not "apt upgrade and forget".
Unsupported versions
The PSAs we fetched name package floors on the current (PVE 9) series only. They do not say "PVE 8 is fine". They also do not give an 8.x version to install. If pveversion still starts with 8., you are outside the numbers in this post - plan the 9 upgrade or get a written backport from Proxmox, do not guess.
FAQ
Is this in CISA KEV? No. Catalog 2026.11.13 / 1,781 entries. No CVE ID exists for 00082 or 00100, so there is nothing for CISA to add under those names.[7]
Is this the TCP UAF kernel PSA? No. That was CVE-2026-98432 (kernel). We already covered it. This post is userspace pve-container / libpve-storage-perl.
Do I need to reboot? Not for 00082/00100 as described. Restart pvedaemon / pveproxy after the debs swap. Reboot only if pveversion -v still shows the old pve-container because the process never exec'd the new files.
Can a CT without a nested mount trick still hit 00100? Different bug. 00100 is upload-error quoting on the node, not the LXC race. Patch both packages.
Should I disable pct mount? The vendor fix is the package bump. Until then, do not run pct mount/pct fstrim as a scheduled job against CTs you do not trust. That is not a substitute for 6.1.17.
Honesty
Vendor-reported, labeled as such: the root-on-host chain for 00082, the root command claim for 00100, the Datastore.AllocateSpace prerequisite for 00088, and "found internally".[1][2][3]
Same dump, same rule: OCI whiteout delete-outside-rootfs, NIC/VNet ACL, ticket impersonation, and the 6.1.15 / 6.1.17 / 9.1.15 floors are the forum posts, not our lab.[4][5][8]
Unverified here: CVSS (none published), CVE IDs (none assigned), CISA exploitation (not in KEV), whether the mount race has been used in the wild, PVE 8 backport versions, how long pve-enterprise lagged pve-no-subscription, and whether 00100 is reachable without a Datastore privilege. We did not reproduce any of these.
Official posts, not the independent mirror: 00082, 00100, 00088, 00094, 00095, 00081.
Sources
[1] https://forum.proxmox.com/posts/874778 - PSA-2026-00082-1 official forum [2] https://forum.proxmox.com/posts/875546 - PSA-2026-00100-1 official forum [3] https://forum.proxmox.com/posts/875369 - PSA-2026-00088-1 official forum [4] https://forum.proxmox.com/posts/875381 - PSA-2026-00094-1 official forum [5] https://forum.proxmox.com/posts/875382 - PSA-2026-00095-1 official forum [6] https://forum.proxmox.com/posts/874776 - PSA-2026-00081-1 official forum [7] https://www.cisa.gov/sites/default/files/feeds/known_exploited_vulnerabilities.json - CISA KEV catalog JSON [8] https://forum.proxmox.com/threads/proxmox-virtual-environment-security-advisories.149331/page-6 - PVE security advisories thread page 6