Currently, many HTML5 elements (e.g. figure, figcaption, aside) are stripped out by HTML Purifier, unless you're a trusted user - i.e. not a student.
Stripping these elements prevents improved semantics in the content, and breaks imported HTML (e.g. imported into Book activities) for some users (whilst leaving it intact for others, which obscures the changes made by HTML Purifier).
EDIT: It turns out that usually the elements aren't stripped in Books, but I was viewing one via log-in-as, which (since MDL-63786) forces text cleaning on for logged-in-as sessions.
Adding HTML5 definitions to HTML Purifier could be done by adding individual rules for each element (like
MDL-54847, which added in support for the audio, video, and source elements), or perhaps by vendoring an existing solution (e.g. https://github.com/xemlock/htmlpurifier-html5).
Support for the semantic elements could be added to Atto later, or via plugins.