On Tue, Apr 12 2022, Jeremy Maitin-Shepard wrote:
> The current fix for CVE-2022-24765 prevents unsafe command execution
> in some cases but does not address all cases:
>
> - Ownership by the current user should not be taken to mean "trusted":
> the user may have retrieved a directory tree from an untrusted source,
> including:
> - Another version control system (which won't prevent a .git directory)
> - Unpacking an archive
> - FUSE mounting a remote filesystem
Those would be good to fix, but I don't think it's correct that it's
within the scope of CVE-2022-24765.
That CVE is specifically about the multi-user case where we'd previously
pick up another user's .git directory.
Whereas these cases aren't like that in that they:
1) Require the user to have set up that .git directory themselves, in
one way or another.
2) ...or for the OS to enforce user permissions so loosely that others
can chown files as you.
> Additionally, the current fix requires additional configuration to
> support existing use cases, and does not provide uses a way to safely
> execute commands like `git status` or `git log` on untrusted
> repositories.
Yes, I agree. I.e. that the method of fixing it is all-or-nothing, and
therefore creates escalation issues that wouldn't occur with a narrower approach.
There was extensive off-list discussion about this, my [1] touches some
of it. To quote my side of a discussion following-up [1] about the more
narrow approach:
An implementation that asks the user to create an opt-in for any and all
config will be much less secure for the core.sharedRepository case,
which is an edge case my proposed change on top doesn't have.
That's because we'll refuse to read *any* config (including just that
created by "git init"), and then ask the user to opt-in to any *future*
config to keep using the repository at all.
So for the sort of staging environment I mentioned upthread it'll
effectively create a vulnerability which we'll still have no practical
mitigation for.
Whereas if we only error on the specific exploitation vector(s) users
who're using a repository without such config will never need to opt-in
at all, so when someone sneaks up on them later and adds a
core.fsmonitor=/path/to/exploit they'll know, and we can show a much
more pointed and obvious error to that effect.
As noted in [1] the solution that got committed does suffer from that
edge case, but I think [1] also summarizes why that approach was taken.
A way of mitigating that case is to put something like this in one's
.bashrc:
git () {
command git -c core.fsmonitor=false "$@";
}
Then, even if you need to use that safe.directory feature you won't be
vulnerable to someone sneaking up on you and adding this to the
.git/config of the (presumably core.sharedRepository) repo:
[core]
fsmonitor = rm -rf /
> I think a better solution would be for git to support a `--safe`
> option that only runs config-specified commands specifically added to
> an allowed list, or if the repository itself has been added to
> safe.directories.
>
> Ideally git would default to running in `--safe` mode, but if that is
> too disruptive at least the option would be available for use in
> prompt commands, etc.
>
> (Please CC me in replies as I'm not subscribed to the list.)
What you're suggesting would be nice, and has been discussed on-list
before.
But it's fundamentally the same case as "make inspecting an unpacked
.git [from a tarball] safe", i.e. that we'll currently pick up config
from it.
The CVE specifically exists because it's subtly different from that
long-known-about case.
1. https://lore.kernel.org/git/220412.86h76yglfe.gmgdl@xxxxxxxxxxxxxxxxxxx/
