2015-10-11-html_white_listed_sanitizer.html

---
layout: post
title:  'A simple yet robust approach to sanitizing user supplied HTML and CSS'
permalink: '/html_white_listed_sanitizer/'
tags: ['html', 'css', 'parser', 'xss']
prettyPrint: true
---
<div class="lead">
  <p>
    Web applications sometimes need to render a piece of HTML
    that has been supplied by the users. This happens for example when
    dealing with content ediable/rich text editors or 3rd party
    integration (ads, games, etc.).
  </p>
  <p>
    The web security risk with having user supplied HTML in a page is obvious:
    if the page fails to properly strip all scripts, a malicious user will be
    able to run arbitrary javascript and hijack the user experience (i.e. the
    page will be vulnerable to XSS).
  </p>
  <p>
    This page presents a a robust way to sanitize user supplied HTML and
    CSS in ~100 lines of JavaScript.
  </p>
</div>

<section>
  <h3>Traditional approaches</h3>
  <p>
    A perfectly safe way to isolate user supplied HTML is to enable a strict
    CSP ruleset, render the content in an iframe or host the entire page
    on a sandbox subdomain.
  </p>
  <p>
    In some cases, these isolation methods aren't flexible enough and web
    developers need a way to sanitize the user supplied HTML. Writing a
    parser works but can be tricky: you need to handle the complexity of
    the HTML specification yet only allow a whitelist of tags.
  </p>
</section>

<section>
  <h3>Leveraging the browser</h3>
  <p>
    We can however let the browser parse the user supplied string (something
    browsers are really good at doing) and then recursively sanitize the DOM
    tree before attaching the content to the page.
  </p>
  <p>
    I believe this approach is very robust for two reasons: we are
    manipulating DOM nodes instead of strings and there is no risk of
    "time of check time to use" bugs because the same browser is used
    to parse the HTML and to render the sanitized string.
  </p>
  <p>
    Given that strings like <code>&lt;sc%00ript&gt;</code> can end up trigger a XSS
    on a subset of browsers, any method which involves parsing user
    supplied strings into a tree and then returning a serialized version
    of a subset of the tree is going to be more bullet-proof than code
    trying to directly sanitize a string. Failing to write a proper parser
    is one of the reasons FBML was riddled with security bugs.
  </p>
  <p>
    Finally I like the approach of walking the DOM tree because it's
    simple and can be implemented in a small amount of JavaScript.
  </p>
</section>

<section>
  <h3>Demo</h3>
  <p>
    Below is a demonstration of this method. The input field lets you enter
    arbitrary HTML but only keeps <code>&lt;a&gt;</code>,
    <code>&lt;img&gt;</code> <code>&lt;p&gt;</code>,
    <code>&lt;div&gt;</code>, <code>&lt;span&gt;</code>,
    <code>&lt;br&gt;</code>, <code>&lt;b&gt;</code>,
    <code>&lt;i&gt;</code> and <code>&lt;u&gt;</code> tags. In addition, the
    sanitizer allows setting <code>border</code>, <code>margin</code> and
    <code>padding</code> CSS properties.
  </p>
  <p>
    Can you get the page to run <code>alert(1)</code>?
  </p>
  <div>
    <textarea id="input" style="width: 800px; height: 100px; border: 1px solid black"><div>
  <b>Lorem</b> <i>Ipsum</i>
  <div style="border: 1px solid green; padding: 5px">
    <a href="http://quaxio.com/html_white_listed_sanitizer/">a link</a> and an image:
    <p><img src="http://quaxio.com/files/2015/html_white_listed_sanitizer/success.jpeg"></p>
  </div>
  <!-- self-closing tags work --> <br/>
  <div style="border: 1px solid red; padding: 5px">
    <a href="javascript:alert(1)">this won't run</a>
    and neither will this: <script>alert(1);</script>
  </div>
</div></textarea>
  </div>
  <button onclick="runSanitizer()">sanitize!</button>
  <div><br/></div>
  <p>Output as string:</p>
  <pre id="output_as_string"></pre>

  <p>Output as html:</p>
  <div id="output_as_node" style="border: 1px solid black; padding: 10px"></div>
</section>

<section>
  <h3>Code</h3>
  <pre id="source_code" class="prettyprint linenums lang-js"></pre>
<script id="html_whitelisted_sanitizer">/**
 * Sanitizer which filters a set of whitelisted tags, attributes and css.
 * For now, the whitelist is small but can be easily extended.
 *
 * @param bool whether to escape or strip undesirable content.
 * @param map of allowed tag-attribute-attribute-parsers.
 * @param array of allowed css elements.
 * @param array of allowed url scheme
 */
function HtmlWhitelistedSanitizer(escape, tags, css, urls) {
  this.escape = escape;
  this.allowedTags = tags;
  this.allowedCss = css;

  // Use the browser to parse the input but create a new HTMLDocument.
  // This won't evaluate any potentially dangerous scripts since the element
  // isn't attached to the window's document. It also won't cause img.src to
  // preload images.
  //
  // To be extra cautious, you can dynamically create an iframe, pass the
  // input to the iframe and get back the sanitized string.
  this.doc = document.implementation.createHTMLDocument();

  if (urls == null) {
    urls = ['http://', 'https://'];
  }

  if (this.allowedTags == null) {
    // Configure small set of default tags
    var unconstrainted = function(x) { return x; };
    var globalAttributes = {
      'dir': unconstrainted,
      'lang': unconstrainted,
      'title': unconstrainted
    };
    var url_sanitizer = HtmlWhitelistedSanitizer.makeUrlSanitizer(urls);
    this.allowedTags = {
      'a': HtmlWhitelistedSanitizer.mergeMap(globalAttributes, {
          'download': unconstrainted,
          'href': url_sanitizer,
          'hreflang': unconstrainted,
          'ping': url_sanitizer,
          'rel': unconstrainted,
          'target': unconstrainted,
          'type': unconstrainted
        }),
      'img': HtmlWhitelistedSanitizer.mergeMap(globalAttributes, {
          'alt': unconstrainted,
          'height': unconstrainted,
          'src': url_sanitizer,
          'width': unconstrainted
        }),
      'p': globalAttributes,
      'div': globalAttributes,
      'span': globalAttributes,
      'br': globalAttributes,
      'b': globalAttributes,
      'i': globalAttributes,
      'u': globalAttributes
    };
  }
  if (this.allowedCss == null) {
    // Small set of default css properties
    this.allowedCss = ['border', 'margin', 'padding'];
  }
}

HtmlWhitelistedSanitizer.makeUrlSanitizer = function(allowed_urls) {
  return function(str) {
    if (!str) { return ''; }
    for (var i in allowed_urls) {
      console.log(allowed_urls[i]);
      if (str.startsWith(allowed_urls[i])) {
        return str;
      }
    }
    return '';
  };
}

HtmlWhitelistedSanitizer.mergeMap = function(/*...*/) {
  var r = {};
  for (var arg in arguments) {
    for (var i in arguments[arg]) {
      r[i] = arguments[arg][i];
    }
  }
  return r;
}

HtmlWhitelistedSanitizer.prototype.sanitizeString = function(input) {
  var div = this.doc.createElement('div');
  div.innerHTML = input;

  // Return the sanitized version of the node.
  return this.sanitizeNode(div).innerHTML;
}

HtmlWhitelistedSanitizer.prototype.sanitizeNode = function(node) {
  // Note: <form> can have it's nodeName overriden by a child node. It's
  // not a big deal here, so we can punt on this.
  var node_name = node.nodeName.toLowerCase();
  if (node_name == '#text') {
    // text nodes are always safe
    return node;
  }
  if (node_name == '#comment') {
    // always strip comments
    return this.doc.createTextNode('');
  }
  if (!this.allowedTags.hasOwnProperty(node_name)) {
    // this node isn't allowed
    console.log("forbidden node: " + node_name);
    if (this.escape) {
      return this.doc.createTextNode(node.outerHTML);
    }
    return this.doc.createTextNode('');
  }

  // create a new node
  var copy = this.doc.createElement(node_name);

  // copy the whitelist of attributes using the per-attribute sanitizer
  for (var n_attr=0; n_attr < node.attributes.length; n_attr++) {
    var attr = node.attributes.item(n_attr).name;
    if (this.allowedTags[node_name].hasOwnProperty(attr)) {
      var sanitizer = this.allowedTags[node_name][attr];
      copy.setAttribute(attr, sanitizer(node.getAttribute(attr)));
    }
  }
  // copy the whitelist of css properties
  for (var css in this.allowedCss) {
    copy.style[this.allowedCss[css]] = node.style[this.allowedCss[css]];
  }

  // recursively sanitize child nodes
  while (node.childNodes.length > 0) {
    var child = node.removeChild(node.childNodes[0]);
    copy.appendChild(this.sanitizeNode(child));
  }
  return copy;
}

function runSanitizer() {
  var parser = new HtmlWhitelistedSanitizer(true);
  var sanitizedHtml = parser.sanitizeString(input.value);
  output_as_string.textContent = sanitizedHtml;
  output_as_node.innerHTML = sanitizedHtml;
}
source_code.innerText = html_whitelisted_sanitizer.innerText
</script>
</section>

<section>
  <h3>Notes, credits and links</h3>
  <p>If you plan to post-process the sanitized DOM, keep in mind that some attribute have side effects which might have already
    taken effect. E.g. setting the src attribute on an image fires an http request right away (even before the image is actually
    added to the DOM). You are probably better off performing all your processing as the new nodes get created.
  </p>
  <p>Credits to Erling for pointing out the need for <code>document.implementation.createHTMLDocument()</code>!</p>
  <p>Credits to Ben Gotow for suggesting some improvements.</p>
  <ul>
    <li><a href="http://theory.stanford.edu/~jcm/papers/esorics09-camera-ready.pdf" class="external">Isolating JavaScript with Filters, Rewriting, and
Wrappers</a></li>
    <li><a href="https://github.com/google/caja" class="external">Google Caja</a></li>
    <li><a href="https://code.google.com/p/es-lab/wiki/SecureEcmaScript" class="external">SES</a></li>
    <li><a href="https://developer.mozilla.org/en-US/docs/Web/HTML/Element/iframe#attr-sandbox" class="external">Sandboxed iframe</a></li>
    <li><a href="https://github.com/cure53/DOMPurify" class="external">DOMPurify</a></li>
  </ul>
</section>