Date: Mon, 12 Aug 2019 17:56:03 -0500
From: John Doe <johndoetouhou@...il.com>
To: fulldisclosure@...lists.org
Subject: [FD] Some interesting facts about gitlab runners

So generally when you create a docker container, you specify what network
you want to create it on right? Well due to historical reasons if you don't
the container is created on the default "bridge0" network.

This network doesn't have service discovery in the proper sense. To have
containers talk to each other by name you need to "link" them, in the
legacy docker sense. But in fact it's possible for any container on this
network, to communicate with any other container on this network, over any
port, whether the containers are linked or not. Sure you need to guess the
IP address but you only have 255 to choose from so it's not that difficult.

The fun fact is that gitlab runners are actually using this bridge0
network. You'd think that each job would create its own network for it and
its services, but as of right now gitlab uses docker's legacy "link"
functionality. Which allows sharing environment variables easily, but
forces them to use the bridge0 network. Using a network per job is actually
one of the many many gitlab TODOs. See here:

https://gitlab.com/gitlab-org/gitlab-runner/issues/1042

Looks like this one was recently backlogged! Oops!

So this means that jobs on a gitlab runner can communicate over the network
with any other job concurrently running on the same gitlab runner. But wait
gitlab jobs don't listen over the network generally right? Well gitlab
services do! https://docs.gitlab.com/ee/ci/services/

Take the following scenario. Alice is running a CI job in a shared gitlab
runner from some place that provides hosted gitlab or something. This job
does the following and is a very common use case:
* Starts a mysql service with the password "p4ssw0rd" (she figures she
doesn't need to use a secure password because it's just a job running
internally, at least mysql's docker image will make you put a password on
your service container unlike redis for instance)
* Downloads a copy of her company's production database from the cache
* Restores it to the mysql service of the gitlab job
* Runs unit tests against it

Meanwhile Bob, who is not an employee at Alice's company creates a gitlab
job that does the following:
* Uses a port scanner to see if there's any containers listening on port
3306 in the IP range of the bridge0 network they're connected to
* If there are attempts to mysql connect to them using username root and a
table of common passwords. Remember the IP of the user isn't restricted
since Bob's IP range is the same as Alice's
* If it's able to connect, do mysqldump | aws s3 cp or something to stream
Alice's internal company data to Bob's S3 compatible storage

If Bob's gitlab job ends up running on the same server as Alice's then
he'll be able to steal her company data.

IMPORTANT NOTES:

gitlab.com is NOT vulnerable, because their shared runners only run one
concurrent job, so Bob's gitlab job will never end up running on the same
server as Alice's.

Your company is probably not vulnerable, because everyone with access to
your internal gitlab goes out to lunch together and gives each other hugs
when they're feeling down and is totally trustworthy.

If you're using some shared runner that allows multiple concurrent jobs
potentially from people you don't trust, and the services in your gitlab
jobs don't require authentication to connect to them, you should probably
stop doing that.

MITIGATION:

Do one or more of the following:

* Don't share gitlab runners with strangers. You should host your own
gitlab runners.
* Use concurrency=1 for your gitlab runners, which is the default, so that
they can only run one job at a time.
* Make sure any gitlab services you use utilize authentication with strong
passwords.

_______________________________________________
Sent through the Full Disclosure mailing list
https://nmap.org/mailman/listinfo/fulldisclosure
Web Archives & RSS: http://seclists.org/fulldisclosure/

Powered by blists - more mailing lists