Friday, 4 January 2019

Chaining 2 low impact bugs into Gitlab RCE. Real World CTF 2018 - Flaglab write-up.


Real World CTF

Recently my team and I went to Zhengzhou, China for Real World CTF event.
Organisation and venue were truly amazing. The format of the competition was a bit different from standard jeopardy-style. There was this not so popular concept of showing exploits on stage. To get points awarded, you had to pop calc or show that you control the device you were hacking into. I really liked that idea. As we all know, those competitions usually are not the most attractive for viewers.
Real World CTF’s approach to this seems to be step in the right direction as far as viewership goes. Exploit presentations is something that viewers can sweat over and cheer for.

As far as challenges go, there were very interesting although very hard, mostly consisted of slightly modified versions of real world applications like vmware, windows.

Some of them even used non-modified versions of real life software, eg. there were tasks with newest router, camera firmware.
I described one of the challenges below.

Chaining 2 low/medium impact bugs into RCE

Although we didn’t solve this task during the CTF (we figured the solution 15 minutes before the end of CTF but didn’t manage to implement in time), the challenge was lots of fun, so I was determined to finish it later on my own. Finally found some time to do so.


Task consisted of only two files:


  image: 'gitlab/gitlab-ce:11.4.7-ce.0'
  restart: always
  hostname: ''
      external_url ''
    - '5080:80'
    - '50443:443'
    - '5022:22'
    - '/srv/gitlab/config:/etc/gitlab'
    - '/srv/gitlab/logs:/var/log/gitlab'
    - '/srv/gitlab/data:/var/opt/gitlab'
    - './steg0_initial_root_password:/steg0_initial_root_password'
    - './flag:/flag:ro'

echo -n \`head -n1337 /dev/urandom | sha512sum | cut -d' ' -f1\` > steg0\_initial\_root_password
Description went something along the lines of: “You may need an 0day”.
So… we had docker-compose that builds gitlab version 11.4.7-ce.0, which at the time of the event was actually second latest, there was some release 5 days before the competition. From docker-compose we can also see that the flag is in filesystem under /flag.

At first we thought the description was just a standard teaser, but after briefly looking at the challenge we knew it may actually be the case.

The only other thing that could be flawed other than gitlab itself is the way of generating initial_root_password. Which unfortunately I don’t believe is the case (at least to my knowledge).


Some time later we stumbled upon gitlab blog post describing security issues fixed in the latest release
There was plenty to choose from, our attention was grabbed by one of them which was reported by employee of Chaitin Tech, organizers of the CTF.

We started to look for any service that is accessible from localhost only and could lead to Arbitrary File Read or RCE.
Such service happens to be redis. This vector was already used in the past, you can read up on it in this gitlab issue.
Redis accepts commands in a line based format. There was only one problem - we couldn’t break lines in Webhooks URLs.
We looked at the blogpost once again.

This looks promising, but it affects different part of code, project mirroring is not Webhooks functionality, or is it?
Thanks to Gitlab being open source software, and the fact that the new version was already released, we knew somewhere in the github repository we would find a patch.
Indeed, we can find both of the bugs fixes.



Looking at the SSRF fix, we realized that the code which checks for loopback/local URLs is actually code placed in shared utils, which Project Mirroring uses as well.

At this point, there were about 15 minutes left of the CTF, we had all the parts needed to create exploit, but just not enough time to implement it.


Knowing all of the above, we can put all the pieces together to create an exploit:

Use SSRF to access redis in Project Mirroring functionality
Finding out what works is as simple as looking at the tests written to check for the bug after the fix was deployed.
  it 'returns true for loopback IPs' do
      expect(described_class.blocked_url?('https://[0:0:0:0:0:ffff:]/foo/foo.git')).to be true
      expect(described_class.blocked_url?('https://[::ffff:]/foo/foo.git')).to be true
      expect(described_class.blocked_url?('https://[::ffff:7f00:1]/foo/foo.git')).to be true
      expect(described_class.blocked_url?('https://[0:0:0:0:0:ffff:]/foo/foo.git')).to be true
      expect(described_class.blocked_url?('https://[::ffff:]/foo/foo.git')).to be true
      expect(described_class.blocked_url?('https://[::ffff:7f00:2]/foo/foo.git')).to be true
Use CRLFs to inject redis commands.
Which is as simple as adding \n to our URL.
  shared_context 'invalid urls' do
        let(:urls_with_CRLF) do
Looking at the issue I mentioned earlier we arrive at payload that looks like this:


sadd resque:gitlab:queues system_hook_push

lpush resque:gitlab:queue:system_hook_push "{\"class\":\"GitlabShellWorker\",\"args\":[\"class_eval\",\"open(\'|whoami > /tmp/a \').read\"],\"retry\":3,\"queue\":\"system_hook_push\",\"jid\":\"4552c3b1225428b18682c901\",\"created_at\":1513714403.8122594,\"enqueued_at\":1513714403.8129568}"


Putting that in mirror functionality results in a 500 returned from gitlab. This happens due to gitlab trying to render our URL and in failing to do so, refusing to respond with anything meaningful.
That’s not helpful given that we still need to trigger the mirror by clicking the little refresh button (or just sending POST to update_now?sync_remote=true)

Doing the latter gives us full RCE.
Exploit can be found over here:
Because there could exist some vulnerable gitlab on the web at the moment of writing this, code has been slightly bugged to prevent skids from using it.

Extracting the flag

That’s where pwning real gitlab system would probably end.

However, Real World CTF network infrastructure routing didn’t allow any connections initiated by gitlab host to LAN or Internet.

To extract the flag, you could use some part of the gitlab interface (icons,popups) to inject flag into it and then read it.


The one thing that I realized solving this challenge is that it’s really hard to estimate what is the real impact of the vulnerability. Multiple vulnerabilities chained can be used to achieve far more than using only one at a time and as far as I know no one keeps track of that.
As presented above, we leveraged two bugs that only chained together matter. This can also work across software, multiples CVE’s with low severity in separate software could result in high severity vulnerability combined.

Does there or should exist database of vulnerabilities that consist of existing ones but combined in some way? Could that enable us to check whole system for higher severity vulnerabilities not just every piece of software individually?