When designing security filters against HTML injection you need to outsmart the attacker, not the browser. HTML’s syntax is more forgiving of mis-nested tags, unterminated elements, and entity-encoding compared to formats like XML. This is a good thing, because it ensures a User-Agent renders a best-effort layout for a web page rather than bailing on errors or typos that would leave visitors staring at blank pages or incomprehensible error messages.

It’s also a bad thing, because User-Agents have to make educated guesses about a page author’s intent when it encounters unexpected markup. This is the kind of situation that leads to browser quirks and inconsistent behavior.

One of HTML5’s improvements is a codified algorithm for parsing content. In the past, browsers not only had quirks, but developers would write content specifically to take advantage of those idiosyncrasies – giving us a world where sites worked with one and only one version of Internet Explorer. A great deal of blame lays at the feet of developers who refused to consider interoperable HTML design and instead chose the principle of Code Relying on Awful Patterns.

Parsing Disharmony

Untidy markup is a security hazard. It makes HTML injection vulns more difficult to detect and block, especially for regex-based countermeasures.

Regular expressions have irregular success as security mechanisms for HTML. While regexes excel at pattern-matching they fare miserably in semantic parsing. Once you start building a state mechanism for element start characters, token delimiters, attribute names, and so on, anything other than a narrowly-focused regex becomes unwieldy at best.

First, let’s take a look at some simple elements with uncommon syntax. Regular readers will recognize a favorite XSS payload of mine, the img tag:

<img/alt=""src="."onerror=alert(9)>

Spaces aren’t required to delimit attribute name/value pairs when the value is marked by quotes. Also, the element name may be separated from its attributes with whitespace or the forward slash. We’re entering forgotten parsing territory. For some sites, this will be a trip to the undiscovered country.

Delimiters are fun to play with. Here’s a case where empty quotes separate the element name from an attribute. Note the difference in value delineation. The id attribute has an unquoted value, so we separate it from the subsequent attribute with a space. The href has an empty value delimited with quotes. The parser doesn’t need whitespace after a quoted value, so we put onclick immediately after.

<a""id=a href=""onclick=alert(9)>foo</a>

Browsers try their best to make sites work. As a consequence, they’ll interpret markup in surprising ways. Here’s an example that mixes start and end tag tokens in order to deliver an XSS payload:

<script/<a>alert(9)</script>

We can adjust the end tag if there’s a filter watching for </script>. In the following payload, note the space between the last </script and </a>.

<script/<a>alert(9)</script </a>

Successful HTML injection thrives on bad mark-up to bypass filters and take advantage of browser quirks. Here’s another case where the browser accepts an incorrectly terminated tag. If the site turns the following payload’s %0d%0a into \r\n (carriage return, line feed) when it places the payload into HTML, then the browser might execute the alert function.

<script%0d%0aalert(9)</script>

Or you might be able to separate the lack of closing > character from the alert function with an intermediate HTML comment:

<script%20<!--%20-->alert(9)</script>

The way browsers deal with whitespace is a notorious source of security problems. The Samy worm exploited IE’s tolerance for splitting a javascript: scheme with a line feed.

<div id=mycode style="BACKGROUND: url('java
script:eval(document.all.mycode.expr)')"
expr="alert(9)"></div>

Or we can throw an entity into the attribute list. The following is bad markup. But if it’s bad markup that bypasses a filter, then it’s a good injection.

<a href=""&amp;/onclick=alert(9)>foo</a>

HTML entities have a special place within parsing and injection attacks. They’re most often used to bypass string-matching. For example, the following three schemes use an entity for the “s” character:

java&#115;cript:alert(9)
java&#x73;cript:alert(9)
java&#x0073;cript:alert(9)

The danger with entities and parsing is that you must keep track of the context in which you decode them. But you also need to keep track of the order in which you resolve entities (or otherwise normalize data) and when you apply security checks. In the previous example, if you had checked for “javascript” in the scheme before resolving the entity, then your filter would have failed. Think of it as a time of check to time of use (TOCTOU) problem that’s affected by data transformation rather than the more commonly thought-of race condition.

Security

User Agents are often forced to second-guess the intended layout of error-ridden pages. HTML5 brings more sanity to parsing markup. But we still don’t have a mechanism to help browsers distinguish between typos, intended behavior, and HTML injection attacks. There’s no equivalent to prepared statements for SQL.

  • Fix the vuln, not the exploit – It’s not uncommon for developers to denylist a string like alert or javascript under the assumption that doing so prevents attacks. That sort of thinking mistakes the payload for the underlying problem. The problem is placing user-supplied data into HTML without taking steps to ensure the browser renders the data as text rather than markup.
  • Test with multiple browsers – A payload that takes advantage of a rendering quirk for browser A isn’t going to exhibit security problems if you’re testing with browser B.
  • Prefer parsing to regex patterns – Regexes may be as effective as they are complex, but you pay a price for complexity. Trying to read someone else’s regex, or even maintaining your own, becomes more error-prone as the pattern becomes longer.
  • Encode characters at their point of presentation – You’ll be more successful at blocking HTML injection attacks if you consistently apply encoding rules for characters like < and > and prevent quotes from breaking attribute values.
  • Define clear expectations – Ambiguity for browsers enables them to recover from errors gracefully. Ambiguity for security weakens the system.

HTML injection attacks try to bypass filters in order to deliver a payload that a browser will render. Security filters should be strict, by not so myopic that they miss “improper” HTML constructs that a browser will happily render.