Isn't a git commit trail basically a Merkle tree of checksums? If any developer tried to do a pull or fetch they'd suddenly get a bunch of strange commit messages, wouldn't they?
Also: code signing is / can become a thing.
I think the release files is the place they could most easily tamper - generally they're stored on Github infra so the files could be changed, and the checksum on the download page also altered (or different files and different checksums provided to different people if targeted).
Unless the builds are totally reproducible it'd be tricky to catch.
To that end, I started a project last month so that code signing can be done in multiple geographical locations at once: https://github.com/soatok/freeon
That Merkle tree prevents the naive case where the adversary tries to serve a version of a repo, to a client who already has an older version, differing in a part the client already has. (The part the client has local checksums for). They shouldn't do that. The git client tells the server what commits it doesn't have, so this is simple to check.
Code signing could be a safeguard if people did it, but here they don't so it's moot. I found no mention of a signing key in this repo's docs.
The checksum tree could be a useful audit if there were a transparency log somewhere that git tools automatically checked against, but there isn't so it's moot. We put full trust in Microsoft's versions.
Lots of things could be helpful, but here and now in front of us is a source tree fully in Microsoft's control, with no visible safeguards against Microsoft doing something evil to it. Just like countless others. It's the default state of trust today.
Git is a distributed vcs after all. Every checkout is its own complete git "hub".
Git may be designed as a distributed VCS; and it'd be a different situation if it were used that way in practice. For many projects, GitHub has a full MITM. They could even—forget about the checksums—bifurcate the views in between devs—accept commits from one dev, send over those commits with translated Merkle trees to another dev who has a corrupted history, and they'd never figure it out.
I think the point is they don't have complete control over it. Sure, they have complete control over the version that is on GitHub. But git is distributed, and the developers will have their own local copies. If Microsoft screwed with the checksums, and git checks them. The next developer pull or push would blow up.
If they're pushing or pulling to/from GitHub, then GitHub has a total MITM and is able to dynamically translate checksum trees in between devs' incompatible views of the repo.
Those checksums would seem valid to the victims, as they're a self-consistent history of checksum trees they got directly from GitHub. The devs would be working with different checksum trees. GitHub would maintain both versions, serving different content and different checksums depending on who asks.
That might work for a while if dev isn't active. He would, for example have to not notice there was a new release, with an incremented version number that triggers updates. Even that doesn't work forever. Down stream dev's often look at the changes - for example a Debian maintainer usually runs his eye over the changes.
But if the dev is active this is going to be noticed pretty quickly. The branches will diverge, commit messages, feature announcements, bug reports, line numbers not matching up. It would require a skilled operator to keep them loosely in sync, and that's the best they could do.
Either way, sooner or later Microsoft's subterfuge would be discovered, and that is the death knell for this scenario. The outrage here and elsewhere would boil over. Open source would leave github en masse, Microsoft's reputation would be destroyed, they would lose top engineers. I don't have a high opinion of Microsoft's technical skills and leadership as they have been consistently demonstrated themselves to be inconsistent and unreliable. But the company too large and too successful to be psychotic. The shareholders, customers, and lawyers would have someones guts for garters if they pulled a stunt like that.
But it's written in rust.
That won't work. The first thing the client does is ask the server for list of references with their oids (ls-refs). It only asks for oids and reports what oids it has after the server responds.
You'd need another way to identify that the client asking for references was the same one you vended the tampered source tree to, otherwise, you'd need to respond with the refs' real oids and the fetch would fail since there's no way to get from the oid the user has to the real one.