Release tarballs and security (xz fallout)

Hi all!

For context: <https://tukaani.org/xz-backdoor/>

Given the recent XZ vulnerability caught just in time, I wonder if we should take any action in this project to help the ecosystem.

Some have mentioned that release tarballs are too opaque, and can easily hide code that's not under git(1) control. I myself have been feeling like that for a long time.

I've modified the build system recently so that tarballs should be reproducible byte-per-byte. This means that downstream distributors don't really need to "trust" tarballs signed by me, but they can (and IMO should) generate them themselves by running make dist, and they should be fine. Our git tags (and all the commits, BTW) are signed.

Here's my proposal:

Stop distributing release tarballs, and instead ask downstream packagers to create them themselves by running make dist, or even not using tarballs at all; make install from a tarball should be exactly the same as make install from the source repository (IIRC).

Any opinions?

Cheers,
Alex

<https://www.alejandro-colomar.es/>

hello, here my opinion ...

I followed the xz hack and the main problem (in my view) is that the person (tried or go) write access to the git archive. In that case it is game over.

Normaly i run a tar vtzf BEFORE i install a tarball (more to make sure that they will produce a directory and do not spill everything in my home). That is be no means perfect but it helps a bit.

For the man-page project i can imagine a 2 tarball solution. One for the pages, and one for anything executeable, so i can use an older/trusted version of an installer, but again an "evil maintainer"-attack is - as any suply-chain-attack - hard to notice and even harder to prevent.

YM2C
wh
________________________________________
Von: Alejandro Colomar <[email protected]>
Gesendet: Dienstag, 9. April 2024 17:29:16
An: [email protected]
Betreff: Release tarballs and security (xz fallout)

Hi all!

For context: <https://tukaani.org/xz-backdoor/>

Given the recent XZ vulnerability caught just in time, I wonder if we should take any action in this project to help the ecosystem.

Some have mentioned that release tarballs are too opaque, and can easily hide code that's not under git(1) control. I myself have been feeling like that for a long time.

I've modified the build system recently so that tarballs should be reproducible byte-per-byte. This means that downstream distributors don't really need to "trust" tarballs signed by me, but they can (and IMO should) generate them themselves by running make dist, and they should be fine. Our git tags (and all the commits, BTW) are signed.

Here's my proposal:

Stop distributing release tarballs, and instead ask downstream packagers to create them themselves by running make dist, or even not using tarballs at all; make install from a tarball should be exactly the same as make install from the source repository (IIRC).

Any opinions?

Cheers,
Alex

--
<https://www.alejandro-colomar.es/>

Hi Walter!

It's nice to hear from you again. :)

Not all of the evil files were under git. Probably because it would have been too suspicious. Some file was injected in the distribution tarball. This is what I'd like to preclude by not having tarballs.

This could be interesting. But the repository is probably simpler to use for everyone. And you can do partial checkouts with something like:

$ git checkout vX.X -- share/mk/
$ git checkout vY.Y -- man*/

so you could use an old build system with newer pages easily (I didn't test the above, so it may need some tweaks, but the idea is there).

The thing is that if I don't create a tarball, there's one fewer step in which an evil maintainer could inject evil stuff.

If downstreams generate their tarballs from the repo, then an evil maintainer has to commit the scripts into the git repository, which can be more easily tracked.

If the tarball is generated by the evil maintainer, then it can inject any evil files without having to do any commit, which would be harder to detect. You'd need to check the file list, which isn't short. In the source repo, we have

	$ find * -type f | wc -l
	2875

You could ignore the pages, and diff the rest to older versions, but that's easier to do if you can use git(1). You can also regenerate the tarball yourself from the sources, then diffoscope(1) it with mine. But at that point, you've already generated a tarball that --if all goes well-- should be identical to mine, so why would you want the official one?

We'd also reduce the need for compressed tarballs, so less dependencies on tools that seem to be dangerous these days.

Have a lovely day!
Alex

<https://www.alejandro-colomar.es/>