Kernel vulnerability CVE-2026-99312

mbosma

Well-Known Member
Dec 3, 2018
126
32
68
32
I just found this CVE in a mailing list regarding the possibility to read memory from host or another vm from inside a vm:
https://lists.openwall.net/linux-cve-announce/2026/11/18/2
I couldn't find anything about the CVE on the forums yet so here I am.

Last update version I saw in the changelogs for the kernel in the enterprise repo had changes up to v7.3.3 and v6.18.62.
The post speaks about 6.18.64 and 7.3.5.

Is the fix stil pending or already implemented in no-subscription?
I'd like to know if the update is already planned so we can plan updating our fleet.
 
Last edited:
Is the fix stil pending or already implemented in no-subscription?
7.3.5 is available in the no-subscription repos. However, 7.3 is still opt-in and, according to this announcement, is not yet recommended for production.

Whether they have already backported the fix to 7.0, and if so, which exact version it was backported to, I don't know unfortunately.
 
Last edited:
I'm getting old. I've read about 10 assessments of this CVE and I'm not 100% certain about the severity, exact attack requirements, or the impact of eventually disabling nesting in kvm_amd (particularily regarding docker in VMs, and chroots for pbuilder and the like).

We do trust the few customers that have root access in VMs (this seems to be required for attacks though I'm not sure), but a combination with other vulnerabilities and hence a chain attack is always a possibility - and access to host RAM can include clear text of private keys and passwords.

So, shouldn't we upgrade and reboot our AMD nodes ASAP?

Shouldn't this be announced in https://forum.proxmox.com/forums/security-advisories.26/ ?

Or, is the attack likely impossible due to some KVM configuration default in PVE?

Or maybe I'm still too naive and VM breakout CVEs are detected and fixed every other day?
 
Of course updating asap is always a good idea, as the idea of a vm is to minimise the risk when hacked. This circumvents that principle.
But... Only when nested is enabled for guests.

More information about nested: https://pve.proxmox.com/wiki/Nested_Virtualization

Docker does not use nesting. As is chroots for pbuilder not a nested vm.

We actually disable nesting on the host. We recommend only using this for testing, or when accepting the risks.
This is not the first cve for nested and it probably won't be the last as the working is: give acces to section in host memory.

I can't speak for proxmox, but read this in another thread on a different cve;
My personal hope is that proxmox will automate the updating process (auto-migrate, restart, check etc)

Hope this helps calm down the nerves :)
 
Hope this helps calm down the nerves :)
Yes it does! Thanks a lot for all the details! :):)

We'll disable nesting too, via sysctl.d/foobar.conf.
And we'll update soon. Due to the recent CVE waves I'm currently focussed on security-hotfixes anyway, documenting and optimizing workflows for hotfixes with deep links, checklists, scripts and the like.

As far as automation goes I'm not so sure.
Last year PVE's HA feature together with overreacting software watchdogs (hard resets that were usually uncalled for) and one node rebooting too slow for HA to re-accept it, we and our customers ended up with data for several VMs and LXCs set back almost a day automatically.
Not Funy.
We've since totally disabled hard resets triggered by the software watch, we don't use any HA automaticsm any more at all, we've migrated some LXCs to VMs (for live migration), and the like. We still use frequent replication (actually way more frequent now), and we do migrate important guests for node reboots (albeit not all, for less important guests it's not always worth the time and resources).

What I'd love to see is live upgrades of the kernel and qemu, but that's not coming anytime soon for any kernel that's not restricted to a small selection of hardware components. There's a reason SuSE gets lots of €s and $s from customers that really really really want to run their biggest machines run 100% 24x7.