OMG, am dying!!!
Someone used the bug GitHub refused to fix, that allows you to add a commit to a repo you don’t control... to upload YouTube-dl to the DMCA request repo on GitHub.
https://github.com/github/dmca/tree/416da574ec0df3388f652e44f7fe71b1e3a4701f
@rysiek I do not, there is a user on hellbird that did this referencing it. I have no knowledge of the reported vuln.
The "bug" is that you can fork someone else's repo, push a commit to your fork, and view the commit on the parent repo, (or anyone else's fork of that repo for that matter) hence the commit hash in the URL. If you click on the commit, you'll see a note that the commit you're viewing isn't part of any branch in that repository.
Take this for example.
https://github.com/thamada/lawn
@rysiek you sir, are disturbingly devious!
@rysiek Perhaps I misread your intent... but perhaps not. :)
@thegibson @Sylvhem not sure if it’s a bug or a feature...
@TheGibson that was @lrvick I think
@TheGibson How to checkout:
git clone https://github.com/github/dmca/
git fetch origin 416da574ec0df3388f652e44f7fe71b1e3a4701f
git checkout 416da574ec0df3388f652e44f7fe71b1e3a4701f
@TheGibson wow, this is just... beautiful :)
@thegibson
It's a cute trick, but this is not a " bug that GitHub refuses to fix" -- it's how git forks work fundamentally. The alternative is to either deep-copy each repo on fork, or to calculate reachability on each direct object request. Both options extremely expensive.
Having two different repos share an object store is a github decision.
@clacke a conscious decision is not a "bug, that GitHub refuses to fix" though. All public forks share objects. Private forks don't.
It is a security vulnerability. Whether or not they decide to fix it.
@thegibson how is it a security vulnerability? The object isn't part of the repo, it's just accessible through an unrelated repo via shared objects. It's confusing, but unless someone copy-pastes from the web interface, it's benign.
Because it creates a culture of weak authorization in ever project, and lets people do things like upload Youtube-dl to a repo that isn't theirs?
I mean ultimately we can record it as a feature if you want, but what if someone decided to just start pumping multiple TB of randomized data into a downstream project?
At the very least his reduces github's ability to meet compliance standards, and is a problem for them....
But as we can see right here, it can be very effectively, and practically abused.
I mean I see models of failure due to this lack of downstream control for the maintainer of the attacked repo.
but this hasn't "uploaded Youtube-dl to a repo that isn't theirs," though. If you clone the DMCA repo, you won't find these objects there, because they aren't part of it. This merely allows someone to view objects via arbitrary web paths. It's confusing -- and can be amusing (https://github.com/torvalds/linux/blob/b4061a10fc29010a610ff2b5b20160d7335e69bf/drivers/hid/hid-samsung.c#L113-L118), but it's not really different than using debugfs to show arbitrary inode contents regardless of where you are in the actual FS tree.
I agree that it would be best if objects not belonging to that specific tree would return a 404, but until git learns to do this cheaply, this is not feasible for projects like github (or git.kernel.org, for that matter) to do, as this would invite a way to do a very simple and efficient DoS via requesting a bunch of arbitrary object paths and forcing the server end to run a bunch of expensive reachability calculations.
I was assuming a long time ago that they didn't even use git repos on-disk, so I'm surprised to see abstractions leak like this.
Well, git protocol doesn't solve this either, it merely refuses to serve you a direct object request unless it's associated with a ref (tag/branch). You can allow requesting only reachable objects via uploadpack.allowReachableSHA1InWant, but even the manpage sternly warns you that it's expensive and doesn't quite solve the problem anyway (https://git-scm.com/docs/git-config#Documentation/git-config.txt-uploadpackallowReachableSHA1InWant).
@monsieuricon @thegibson @notclacke Oh, cool, the DT Overlay configfs interface has been accepted in Linus' tree ;-)
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=184c79e4acf5a55f1d6febcf4cf1b545880c2fde
Right, kernel.org is susceptible, too..
@geert This is how we only need 18G for all of git.kernel.org instead of 300G. :)
@monsieuricon @thegibson @notclacke I can put there whatever I want ;-)
If I would use a different name/email, could you find out who brought in the commit into kernel.org?
@thegibson That is just amazingly hilarious.