Re: [PATCH v2 2/2] setup: allow Git.pm to do unsafe repo checking

From: Jeff King <peff@peff.net>
Date: Oct 22, 2022, 20:55 UTC
Message-ID: <Y1RY38RULkqd9pBN@coredump.intra.peff.net>
In-Reply-To: <221022.86eduzeiek.gmgdl@evledraar.gmail.com>

On Sat, Oct 22, 2022 at 09:45:14PM +0200, Ævar Arnfjörð Bjarmason wrote:

Show 17 quoted lines

> The vulnerability safe.directory was supposed to address was one where
> you'd set your fsmonitor hook via a config variable, so running "diff",
> "status" etc. would unexpectedly execute arbitrary code.
> 
> Especially on Windows where apparently the equivalent of the root of a
> shared mounted volume routinely has global write permissions (user's
> subdirectories being less permissive).
> 
> An alternative I raised on the security list was to narrowly target just
> the fsmonitor config, since that was the vulnerability.
>
> [...]
>
> I'm unaware of any other variable(s) that provide a similar vector, and
> safe.directory is encouraging users (especially in core.sharedRepository
> settings) to mark a dir as "safe", and we'd then later have an exploit
> from a user with rw access who'd use the fsmonitor hook vector.

Here are a few off the top of my head that you can trigger via git-diff:

  - core.pager will run an arbitrary program

  - pager.diff will run an arbitrary program

  - diff.*.textconv runs an arbitrary program; you need matching
    .gitattributes, but those are under the control of the repository.
    (not diff.*.command, though, as you need to pass --ext-diff)

  - browser/man paths if you run "git diff --help"

And of course as you expand the set of commands there are more options. E.g., credential helper commands if you do anything that wants auth, interactive diff-filter for "add -p", hooks for git-commit, git-push, etc. Those commands are less likely to be run in an untrusted repository than inspection commands like "status" or "diff", but the boundary is getting quite fuzzy.

fsmonitor _might_ be the only one that is triggered by git-prompt.sh, because it has a limited command palette, generally reads (or sends to /dev/null) the stdout of commands (preventing pager invocation), and doesn't do text diffs (so no textconv). Even if true, I'm not sure if that's a good place to draw the line, though. People do tend to run "git log" themselves.

-Peff