GitHub RCE by Environment variable injection Bug Bounty writeup

Disclaimer: I’ll keep this really short but I hope you’ll get the key points.

GitHub blogged a while ago about some internal tool called gerve: https://github.com/blog/530-how-we-made-github-fast

Upon git+sshing to github.com gerve basically looks up your permission on the repo you want to interact with. Then it bounces you further in another forced SSH session to the back end where the repo actually is.

At some point I figured that it is possible to inject some environment variables into gerve/the forked SSH process by setting my username to something like joerchen\n\nLD_ASSUME_KERNEL=1\n\n.

LD_ASSUME_KERNEL=1 will prevent the actual command from being run, just like this:

joernchen ~ $ LD_ASSUME_KERNEL=1 uname -a
uname: error while loading shared libraries: libc.so.6: cannot open shared object file: No such file or directory

For the details on this, check man 8 ld.so.

So far so good, how can we use this fact to make SSH execute arbitrary commands?

The technique I came up with used both features of ld.so and SSH itself:

LD_PRELOAD=/path/to/libfakeroot.so
SSH_ASKPASS=/usr/bin/ex
DISPLAY=:1

How and why did this work?

libfakeroot makes SSH think it’s root (we can inject this via LD_PRELOAD because the ssh binary is not setuid)
ssh tries to read /root/.ssh/known_hosts
ssh fails reading ‘cause it’s actually running as the git user
ssh connects to $backend and wants to ask the user if $backend_hostkey is OK.
ssh has no terminal and DISPLAY is set
ssh invokes the command specified in SSH_ASKPASS

From being dropped in /usr/bin/ex we could just say: !/bin/sh and be happy with having a shell as git@github.com