Examples in the spec seem to suggest inserting HTML markup into HTML's alt and title attributes.
I think that's a mistake. HTML defines these attributes to contain plain text, so a screen reader will announce ![*this*]() as some variation of:
less than em greater than this less than slash em greater than
And  will render as <em>this</em> in the tooltip (where tooltips are supported).
This gives readers text that is completely different than what the author wrote. It's not something anybody would like to do, it exposes an implementation quirk.
I think formatting should stay allowed in Markdown's image alt/title syntax (so conversion straight to something more expressive than HTML can take advantage of that), but Markdown-to-HTML converters must be required to handle limitation of HTML attributes gracefully, without throwing unparsed angle brackets at users.
The simplest rule may be: parse alt and title as Markdown, strip all HTML tags, HTML-escape the result where necessary:
![*foo*]() <img alt="foo">
![\*foo\\*]() <img alt="*foo\*">
![& & "]() <img alt="& & "">
![<div>1 < 2</div>]() <img alt="1 < 2">
