It is a great idea but the concept is flawed

This issue helps to understand it

seccomp background

libseccomp is a long established security “firewall for syscalls”, it allows to restrict what actions programs can perform on the system.

I dont know what exactly these are, often low level stuff I guess, but one of these actions is “create an unprivileged user namespace”.

Flatpak uses a single, “badness enumerating” seccomp filter, which means they block the syscalls “A”, “B” and “C” for all programs. All others are allowed, and (see the issue) programs cannot define a more restrictive one.

user namespaces

A namespace is a virtual filesystem on top of the “real” one, where certain real system files are mounted to. This means unprivileged programs can suddenly do “OS stuff”, needed to create this filesystem.

contra user namespaces

Many distros like Debian (and Arch) didnt enable this feature for a long time, because user namespaces mean that unprivileged userspace programs, like your browser, can suddenly access low level system components like the filesystem.

If there now is a bug in such a component, user namespaces allow the program to directly access a way more privileged area, escape here and thus have privilege escalation.

This would not be possible when not using user namespaces.

Flatpak has a seccomp filter that blocks the creation of user namespaces, to avoid this low level system access. Which is a very good thing, but the missing modularity doesnt allow anything else!

pro usernamespaces

Over time even slow pacing distros like Debian enabled user namespaces.

Today they are the core feature of bubblewrap, podman, docker, (and thus distrobox and toolbox)…

The concept is that a rootless binary, running in userspace, can access system components which would normally require root.

Firejail is the opposite example, it is a root binary and sandboxes apps also when user namespaces are disabled. Chromium has a fallback suid sandbox which also is a root, same with bubblewrap after the modification by 34N0 implemented in the “no userns” images of secureblue.

The problem is, a root binary has root access. If there is a flaw in it, which was the case with firejail, the “nice and secure isolated app” could now use the root binary to escalate its privileges to root level, more than what it could have done without it.

Browsers and Sandboxes

A browser is basically a platform to run “apps” on. Nearly all websites nowadays require executable code, which means browsers are the attack surface for malware. Scrap your verified Flathub or well maintained distro repository, a single website could use a weakness and break your system.

This was a thing back in the time… crazy huh?

Chromium (Chrome, Edge, Brave, Vivaldi, Opera, …) has the said rootful sandbox as a fallback, I guess implemented back when user namespace sandboxes were not adopted enough.

But is normally uses a user namespace sandbox for process isolation, every tab runs in a different process, on Android too.

Firefox also uses user namespace sandboxes for tabs, but additionally uses seccomp-bpf to restrict the syscalls that the isolated tabs/processes can execute.

Flatpak and Chromium

Chromium relies exclusively on the ability to create sandboxes, with a root binary (the strange not really used fallback method) or with user namespaces.

So much that it straight up doesnt run if it cannot do that. The same goes for Electron apps, which are a browser platform running a single or very few processes.

This is why zypak was created. It redirects the calls of Chromium to flatpak, so it uses the builtin Flatpak sandbox instead.

As I said, all Flatpak apps (and thus all processes) use the same seccomp filter, so I assume that zypak is less secure than the native sandbox, which is battle tested by Google, Microsoft and more companies.

But it uses a sandbox, it is rootless and uses user namespaces. It just needs a little testing, a security audit, a bit of pentesting.

At the current stage I would honestly not trust it, so Flatpak Chromium browsers are not recommended for “production”.

Flatpak and Firefox

The Firefox Flatpak is official.

What the heck?

Thats what I asked byself and created this issue but honestly, this issue thread is way more informative.

Right now we just fork(), so replacing that with flatpak-spawn would cause a massive increase in memory usage? You would no longer have CoW sharing of memory.

So Firefox would need big architectural changes to support a sandbox like Flatpak’s. It uses copy-on-write to save Memory and be more efficient.

For some reason Chromium works just fine with zypak.

it’s not clear to me the “Flatpak Sandbox” it’s creating is comparable to what we have now (even with just seccomp-bpf). We launch our subprocesses with specific, nailed down sandboxes.

They should absolutely compare their seccomp filters. But this indicates the same issue as the one at the beginning, always using the same seccomp filter is not suited for an entire platform like a browser.

Fair usage in Flatpak

To sum it up:

Electron apps are likely fine to be ran as Flatpaks. The zypak sandbox may not isolate the processes from another as well as the normal one does, but they are controlled and known code.

Electron uses Chromium because of laziness, not because it needs the security of the platform. Daniel Micay, the creator of GrapheneOS, would also list a few very technical things why Electron has crippled security features of Chromium.

Thunderbird is using Firefox similar to Electron, just as a platform for known code, so this will be fair too.

Flatpak Firefox… is probably okay secure. If you use UBlock Origin with some filterlists, and an opt-in NoScript setup (which I highly recommend for privacy and security), the risk is even lower.

But the risk is literally getting malware, losing all your data, getting breached or intruded. So why leave out this security measurement.

But, its true, Flatpak isolates the browsers from the system, which is really nice. If there is a weakness in the browser platform, a process could not just escalate and access everything Firefox can.

Bubblejail

So isolating the browser from the system using Bubblewrap, a modern and rootless sandboxing tool, sounds like a good idea.

The only issue is the always-the-same seccomp filter. The best solution would be a fix for the issue at the beginning, but for now we can use bubblejail.

It is a tool that makes the creation of bubblewrap and seccomp filters easy, and adds Desktop entries to launch existing apps through that sandbox.

For some reason it doesnt work at all anymore for me… but it did in the past. It is certainly not ready, but with some helping hands it can fix all the gaps, where system apps are needed for certain abilities.

May that me a VPN app, Nextcloud-client adding icons to your task manager, an IDE like VSCodium, Zed, Lapce, Kate… or isolating all your system apps!

---

So currently I use Fedora Firefox, which is very well maintained and checked for security build flags.

I will continue making bubblejail work, which will be a good solution for this problem.

Distrobox will introduce a startup lag of it’s own. I would rather recommend (seriously this time) something like nix (it is officially supported for your distro) or junest.

You can try rpm-ostree to install in the lowest userspace layer, but Bazzite purposely switched to flatpak Firefox and discourages installing software using rpm-ostree, unless you absolutely have to.

While toolbox and distrobox seem very similar, distrobox comes with a slight warning :

• https://distrobox.it/#security-implications
• https://wiki.archlinux.org/title/Distrobox#Security_implications