Hacker News Reader

> Meanwhile here I am on my Linux machine, constantly anxious that sooner or later one of my bazillion npm and pip dependencies will get compromised, and secretly praying that one day proper sandboxing and an Android-security model will be common on the Linux desktop, so that I can erect security boundaries between my applications and repositories.

Why wait? You can shove your pip/npm uses into docker/podman and remove 90% of the attack surface today. (Provided you don't map your home directory into the containers)

Docker is not a security barrier. There have been plenty of container escape attacks in the past, and plenty more to come.

But I agree it might remove the 90%.

No, Linux containers are a security barrier. They're not absolutely 100% impenetrable, but they're pretty good, and I would generally expect them to be fine against the kind of things most people are worried about getting from pip/npm.

Firejail and apparmor have existed for years. If you don't use them maybe it's your fault?

Also the very same npm backdoors have already hit android apps. What can sandboxing do if you backdoor a dependency of your banking app?

Or go old-school with multiple users and chroots? You could even install from (and host) a trusted repository, where the source and binaries are vetted (and you can pay people to do this for you).

Server software is usually compartmentalized in uid:s but desktop software seldom is, if ever. Package managers and maintainers could do a lot here to make it easier. Some things long time Linux users like to do, like running Firefox as a separate user, is still a much more involved process than it should be.

A lot of it is probably standards and culture work, like where a user can expect to store files and have them readable by Firefox in this example. So perhaps this is something the GNOME/Freedesktop people could have been interested in and made a difference? Instead we have things like Flatpak, which is good but not the lowest hanging fruit here.

For user-facing stuff, I agree it's hard because of the challenge of managing access to data (and I would argue no system does this well, Android has a different set of failure modes, and I've not used QubesOS but presumably it has it's own issues as well), but in the top-level comment, the concern was around using pip/npm, which to me is almost a solved problem if you care enough and are willing to put the effort (and money) in.

It's also not like Linux is any different with respect to installing random PyPI/npm packages on any other desktop/laptop OS (https://xkcd.com/1200/), so I'm not sure anything desktop Linux does here would change the fact that installing random software from the internet may be a bad idea sometimes ;)

Completely agreed on this. Linux, by and large, should actually be far easier here? Have a "work account" for your machine where you do these tasks and you are basically there. Switching to a gaming account or your banking/etc. seems easy enough?

You're going to deal with the users who can't attach a file to an email because the firefox process has no access to it?

To be fair, if firefox had the intelligence to know that it was being asked to attach a file it didn't have access to, it could prompt for a password. I don't expect full TRAMP like smarts from Emacs, but I don't see why this wouldn't be doable?

Granted, I'm viewing this as far easier than the sandbox "fake file system" approach? Firefox would be able to see the file exists, most likely, but just not have read rights to it. Yes, you can have some things it can't list, but I would expect that to be low on probability to want to attach to an email?

The Flatpak/XDG people are standardizing the concept of "portal"s for doing things that need help from outside the sandbox. A properly sandboxed Flatpak app can only open files bind-mounted for it via a portal call.

That's exactly my point. Users expect desktop software to read their home directory, which is an artefact of running under the same uid. The same can not be said for server software, where the expectation is different.

Linux desktop environments had the chance to set a precedent here where software could for example have had a directory named after the application under the user's home, and use subuid to access files. That is simply an example, but would be a backwards compatible way to create a shared culture how Linux desktop software operate, and easen a transition where desktop software could be limited to "their" own directory.

Instead they squandered this chance by focusing all effort on moving dotfiles around (.config was such a waste of energy where we the heated debate could have been about something more useful, such as the above). Now we have Snap and Flatpak which both try to solve a culture problem (where we store our files) with technical solutions (bind mounts! modal gui:s! popup dialogs!). These will not improve the situation. In the best case it will train users to click "Yes" more, which we know not to help security.

Sandboxing should be built in and by default, not DIY and glued on, like with apparmor and firejail.

"Your car does not come with a seatbelt? Seatbelt parts are easy to order online and assembled on any car, it's your fault for not using one."

> Also the very same npm backdoors have already hit android apps. What can sandboxing do if you backdoor a dependency of your banking app?

The whole point of sandboxing is that one compromised app can not compromise the whole system and other apps. Compromised dependency on my banking app on Android or iOS only compromises that banking app and nothing else.

Fedora Silverblue is this

It may be in the future, but for now it is no different from Fedora Workstation in terms of security. Please correct me if I am wrong. AFAIK Silverblue has no additional sandboxing or any other improvements to security.

Pretty sure Fedora, being based on Red Hat, has the strongest SELinux policy in place by default, and SELinux is pretty much the best sandboxing option available other than actual virtualization.

Yes, but this was about Silverblue and how it implements some additional sandboxing, which it doesn't. SELinux is great, but maintaining it and creating configs is huge amount of work and where on AOSP, every process is strictly confined with SELinux, on Fedora, not so much. Not to mention the additional software the user installs. Not at all comparable to real Android or iOS sandboxing.

Secureblue has entered the chat.

It's generally only initial work to make the policies to maintain a program, maintaining doesn't even really exist unless the program radically changes in some way.

Fedora is notable because any software installed via repositories has a policy written for it, so it is already far more in effect than you might realize.

It's entirely comparable to Android sandboxing because it's part of the foundation of Android sandboxing.

How so? I'm writing this from an Fedora Sericea, which is Silverblue but with Sway instead of GNOME. Atomic Fedoras solve only package hysteresis (your package manager being unable to reproduce the intended system state because of unaccounted for changes) by generating the root file system with OSTree. It has nothing to do with sandboxing the applications themselves.

It does in the sense that all the applications you install will be via flatpak, so they get sandboxed that way. Of course it depends on how locked down the sandbox is configured for each of those applications.

The S in Flatpak stands for Security.

Flatpak is primarily a convenience mechanism for app makers. Any security boundary you may find in it is optional, all defaults are always toward not breaking apps. Apps pretty much uniformly either silently get read access to all your files, and even when that is not true they often get permanent read-write access to any file you open in them.

Go look at the permissions for GNOME Papers. Try to argue that it's "sandboxed".

>Apps pretty much uniformly either silently get read access to all your files

This is outdated information. The situation has improved since the publishing of flatkill with flathub loudly warning about permissions and less apps having full R/W access.

Android apps can be configured insecurely too although less severe, still it's the users responsibility to check and modify permission.

In either case it's a substantial improvement from no isolation at all with much easier handling than other sanbox tools or MACs.

> less apps having full R/W access.

Not good enough when apps can still silently have full access to home and /media without the user even realizing.

It’s always felt strange that Linux desktops try to make sandboxing and permissions the responsibility of packaging standards. That strikes me as much more of a system level thing like audio or display output.

- opensnitch

- flatpak

- docker/podman/vm, etc

- /etc/shadow has been around for decades.

- Boot/login with TPM / Yubikey etc around for a decade.

You can have Qubes OS on the certain PCs, in the same way you can have GrapheneOS on a certain family of the safe hardware.

This is a solved problem if you trust the packaging folks for your distro. Most end-users will never need to install some random stuff from npm or pypi: these are developer-specific concerns.

> Most end-users [...] these are developer-specific concerns

I'd wager a bet and say most end-users who end up using Linux are, by one definition or more, developers.

Just another reason not to needlessly update dependencies. To say nothing about the risk of compromising legacy code. And if you are someone who updates your dependencies constantly just because, consider that for many of the packages you are updating into they don't even do that and use some ancient dependency themselves owing to legacy code issue and the fact everyone for some reason wants to rename all their functions and flags every major version change.