The XML FAQ

You should almost never need to use CDATA Sections. The CDATA mechanism was designed to let an author quote fragments of text containing markup characters (the open-angle-bracket and the ampersand), for example when documenting XML (this FAQ uses CDATA Sections quite a lot, for obvious reasons). A CDATA Section turns off markup recognition for the duration of the section (it gets turned on again only by the closing sequence of double end-square-brackets and a close-angle-bracket).

Consequently, nothing in a CDATA section can ever be recognised as anything to do with markup: it's just a string of opaque characters, and if you use an XML transformation language like XSLT, any markup characters in it will get turned into their character entity equivalents.

If you try, for example, to use:

some text with <![CDATA[<em>markup</em>]]&gt; in it.
	

in the expectation that the embedded markup would remain untouched, it won't: it will just output

some text with <em>markup</em> in it.
	

In other words, CDATA Sections cannot preserve the embedded markup as markup. Normally this is exactly what you want because this technique was designed to let people do things like write documentation about markup. It was not designed to allow the passing of little chunks of bogus, corrupt, or invalid unparsed HTML embedded inside your own XML through to a subsequent process — because that would risk invalidating the output.

As a result you cannot expect to keep markup untouched simply because it looked as if it was safely ‘hidden’ inside a CDATA section: it can't be used as a magic shield to preserve HTML markup for future use as markup, only as characters.

Read How can I handle embedded HTML in my XML? as well, which is very closely related.