Skip to content
Branch: master
Find file History
Latest commit fbcd330 Jun 28, 2019

Google CTF 2019 Quals: Paste-Tastic! (web) writeup

This was one of the 4 web challenges on the CTF, but this was the only web one which had 0 solves when the clock hit at the end of the CTF.

I've just missed the deadline of solving the challenge by ~30 minutes.

(This would've meant 3rd place with the +500pts, and my teammate almost finished an another challenge, so even 1st place would've been possible. We finished 7th in the end.)

So in overall I liked this challenge the most on GCTF (from those I worked with), and although we (!SpamAndHex) qualified to the Finals (the third time actually :P) which made me super happy, still I finished the CTF with a bittersweet feeling of not finishing this challenge on time...

I'd like to thank the Google guys these nice challenges and the superb CTF (Quals + Finals too), definitely one of (if not) the best CTFs in the year!

The challenge

The description of the challenge wasn't too talkative:

The pretty paste solution!

The website had basically 3 important functionality:

  • create a paste
  • view a paste
  • report a paste to the admin

From the "report something to the admin" functionality it was clear that our goal is to achieve XSS on the website. It used Recaptcha protection for the submission.

You could also select your paste's language from a long list of 91 languages and depending your choice your paste was rendered with

  • Marked.js, a Markdown-to-HTML converter, if you chose Markdown
  • Prism, a code syntax highlighter if you chose something from the other 90 options

The app's client-side functionality was implemented in the app.js file.

(The app also remembered your recent 10 pastes and stored them in local storage, but that functionality does not matter for the challenge.)

The plugin loader / dependency system

The app loaded only the necessary plugins depending on what type of paste you created.

For example this was used if your paste was a Markdown file:

  "viewer": [{
    "dependencies": [{
      }, {
        "integrity":"sha256-Y0YX22e5n0zVSAd1tJ6aypkv9o4AEX5YcRKPg1Al8jg=" } ],
    "plugins": {
      "markdown": {
        "requires": ["markup"] },
      "markup": {
        "integrity":"sha256-8nT1E50WC5TDeb3+USsFEXN5ZGgLdmwZ6RS5KT71Wjs=" } },
    "preload": ["markdown"] }]};

The renderer sandbox (iframe)

If you viewed a paste, then the app rendered the content client-side (and not server-side) with Prism or Marked.js, but the rendered content was put into a sandboxed iframe, and not directly into the website's content:

    this.sandbox = document.createElement('iframe');
    this.sandbox.setAttribute('src', '/sandbox');
    this.sandbox.setAttribute('sandbox', 'allow-same-origin');
    this.sandbox.setAttribute('class', 'sandbox');

The sandbox attribute limited what this iframe can do: for example it did NOT allow to run scripts inside the iframe (as allow-scripts attribute was missing).

The allow-same-origin attribute was required, so the website could modify its contents, resize it, scroll into it, etc.

The sandbox iframe and the website communicated via postMessage mechanism.

The Content-Security-Policy (CSP)

The website was also protected by the following Content-Security-Policy:

script-src 'nonce-3ra+TpXGQImBZW8NNdCJ2A==' 'unsafe-eval' 'strict-dynamic' https: http:; base-uri 'none'; object-src 'none'

This means the following:

  • script-src
    • 'nonce-3ra+TpXGQImBZW8NNdCJ2A==': inline scripts can only run if they have a nonce attribute with the value 3ra+TpXGQImBZW8NNdCJ2A==
      • this is how the CONFIG variable was set via <script nonce="3ra...">CONFIG=...</script>
      • (note: the exact value changed with every page load, so was not predictable by the attacker)
    • 'unsafe-eval': eval could be called and eval was actually used by app.js to load dependency scripts
    • 'strict-dynamic': if an otherwise trusted script loads another script then trust shell be propagated to that script too
    • https: http:: scripts could be loaded from HTTP and HTTPS sources
  • base-uri 'none': you could not set base URL via the <base> element (does not really matter for us)
  • object-src 'none': you could not load content via the <object>, <embed> and <applet> elements (does not really matter for us)

The solution

Let's go through the building blocks required to solve this challenge.

Some initial observations

  • There were no (known) vulnerability in the third-party libraries (Prism and Marked.js)
    • edit: turns out Marked's sanitizer can be bypassed with e.g. <script><im<script>g onerror=alert(1) /> as this will be rendered as <img onerror=alert(1) /> (seen in @LiveOverflow's video)
    • I've sent a PR to Marked which deprecates the sanitize argument and recommends using an external sanitizer library, e.g. DOMPurify. Also tries to fix the current bypass (while accepting the fact that other bypasses probably still exist).
  • The admin visited any reported URL, not just the posts on Pastetastic
    • you had to modify the "DMCA" report's HTTP request after the Recaptcha check
  • The Marked library used for rendering Markdown is able to include <img> tag with src attribute and adds id attribute to heading tags (eg. to # -> h1)
    • # ![someName](URL) will be converted to <h1 id="someName"><img src="URL" ...></h1>
    • here is the code which calculates the id value (requires headerIds to be set to true, but this is the default config)

Vulnerability: there is no origin check for one of the postMessage handling code

  • The website's postMessage handling code (app.js:13) did not check the origin and could be called by untrusted parties (us :P)

Abusing Chrome's XSS Auditor

  • CONFIG JS variable can be removed by abusing Chrome's XSS auditor via sending <script ...>CONFIG=...</script> as query parameter (e.g. ?PlzRemoveThisCodeForMe=<script>...)
    • this is possible since Chrome 74 as it switched to filter mode by default instead of previous block
    • node is removed even if the nonce attribute does not match, don't exactly know if it is intentional or not...

Abusing cross-origin frame manipulation

  • CONFIG variable can be reintroduced by iframeing the website and modifying the name of one of its internal iframes
    • there is no X-Frame-Options: DENY header and no frame-ancestors policy in the CSP, so iframeing is possible
    • to modify the name of the iframe to CONFIG, you first have to set the location of the iframe to a same-origin location (same-origin as top frame)
    • iframes are available on window as properties by their name and window properties are visible the same way as global JS variables - read more about in the specs

Abusing the fact that Recaptcha adds (non-sandboxed) iframes into websites

  • You cannot run scripts in the sandbox iframe (as its sandbox attribute does not include allow-scripts value), but Recaptcha also adds two iframes which do not have the sandbox attribute set, so they can run scripts
    • so redirect one of Recaptcha's iframe (not the sandbox one) and name it as CONFIG

Hijacking application's code via DOM Clobbering

  • By creating a fake DOM structure (via DOM Clobbering) with the help of the tricks described above, it is possible to hijack app.js's code and make it to call loadScripts with our JS file (which will eval our code and steal the flag in the cookie)
    • this.config = CONFIG.viewer[index]; (app.js:24)
      • CONFIG is our hijacked Recaptcha iframe pointing to our html file
      • our html file contains <iframe name="viewer" src=""></iframe> which points to a Markdown file
        • this would not work if we used the sandbox iframe, as rendering Markdown requires JS, that's why we used one of Recaptcha's iframe
      • CONFIG.viewer is the Markdown viewer (/view/) page's window
      • CONFIG.viewer[index] (index=0) is the Markdown viewer page's first iframe (so the sandbox iframe) and this is where the rendered Markdown is
      • so this.config is our rendered Markdown's window
    • for loop on this.config.dependencies (app.js:28)
      • Markdown "# dependencies" was converted to <h1 id="dependencies">...</h1>
      • named elements are visible on the window (aka. on this.config) - again: here is the specs
      • so this.config.dependencies is our <h1> tag (in JS it's a [object HTMLHeadingElement])
      • this.config.dependencies.length is undefined as HTMLHeadingElement does not have a length property
        • we had to put dependencies into Markdown, because otherwise this code would've break and stop the script from running (as if dependencies was undefined then dependencies.length would cause a "cannot access a property of undefined" error)
      • i < this.config.dependencies.length is false as 0 < undefined == false in JS
      • so the for loop never runs
    • for loop on this.config.preload (app.js:33)
      • same as before for dependencies
    • this.loadPlugin( (app.js:16)
      • can be anything and is fully controlled by us via postMessage call
    • const spec = this.config.plugins[lang]; (app.js:53)
      • Markdown "# ![plugins](" was converted to <h1 id="plugins"><img src="" ...></h1>
      • this.config.plugins is the <h1> tag
      • set lang to firstElementChild
      • this.config.plugins[lang] is h1's firstElementChild, so the <img> tag
      • spec.requires (app.js:54) is undefined as HTMLImageElement does not have an requires property, so this if is ignored
    • this.loadScript(spec); (app.js:59)
      • so spec is the <img> tag (HTMLImageElement)
      • spec.integrity (app.js:68) is undefined as HTMLImageElement does not have an integrity property, so this if is ignored
      • fetch(spec.src, params) (app.js:71) fetches the <img> tag's src attribute which is our exploit's URL
        • CORS should be considered, so the JS file should be served with the following headers:
          • Access-Control-Allow-Origin: *
          • Access-Control-Allow-Methods: GET

You needed to use HTTPS to access to flag

  • Flag in admin's browser was only available via HTTPS (secure flag in set on cookie)
    • so use HTTPS website for each step, not to run into mixed content errors

The flag


TADA! ;)

You can’t perform that action at this time.