HTML5 specification parser differential

Observations

We have noticed a number of parsing differences between the Masterminds/html5-php parser and the HTML5 specification. We think that the root cause of those issues drills down to the use of PHP’s default parser, loadHTML, DOMImplementation, etc. The lack of HTML5 support by PHP is known and we contacted them asking to make it more clear in the documentation in order to raise awareness for these security issues.
This behavior becomes security-relevant when HTML sanitizers use the Masterminds/html5-php parser. We have come across multiple PHP sanitizers that are vulnerable to bypasses due to using Masterminds/html5-php.

Exploitation

Here are examples of the differentials, and how attackers can leverage these in order to bypass sanitizers.

Comments:

According to the XML specification (XHTML), comments must end with the characters -->.
On the other hand, the HTML specification states that a comment's text 'must not start with the string >, nor start with the string ->'.
When parsing the following string in a browser, the comment will end before the p tag. But when parsing with Masterminds/html5-php the p tag will be considered a comment:

• Input: <!---><p>
• Browser (HTML5 specification) output: <!----><p></p>
• Masterminds/html5-php parser output: <!---><p>-->

An attacker can input the following payload <!---><xss>-->. While the parser considers the xss tag as a comment, the browser will end the comment right before and render the xss tag as expected.

Processing instructions (PI) elements (known, but we encounter sanitizer bypasses due to this)

Processing instructions elements exist in XML specification but in HTML5 the characters <? opens a comment and ends it at the first occurrence of greater than >.

Attackers can create the following Processing Instruction <?xml >s<img src=x onerror=alert(1)> ?> and while no img tag is rendered in Masterminds/html5-php the browser will create a comment and end it at the first > character, rendering the img tag.

Foreign content elements

HTML5 introduced two foreign elements (math and svg) which follow different parsing specifications than HTML. Masterminds/html5-php doesn’t take it into account, causing other parsing differentials and sanitizers bypass such as:

• <svg><p><style><!--</style><xss>--></style>

noscript element

Depending if scripting is enabled (enabled by default in browsers) the noscript element parses its content differently:

• If scripting is enabled, then the content is rendered as raw data
• If scripting is disabled, then the content is rendered as HTML

Masterminds/html5-php parses according to disabled scripting, which is different than the default browsers’ parsing.
This is not wrong per se, but still can cause some mXSS such as:
<noscript><p alt="</noscript><img src=x onerror=alert(1)>">

Start with src/HTML5/Parser/DOMTreeBuilder.php, especially the use of PHP's DOMDocument.loadHTML and DOMImplementation. Reproduce the comment, processing-instruction, foreign-content, and noscript differentials against browser HTML5 parsing. Before coding, define the intended scope and acceptance tests for parser behavior and sanitizer safety; the issue does not identify a single fix.