I think your thesis is generally sound, but is too hand wavy at the idea that an ai could be trained, maligned, escape and then hide itself à la Robopocalypse by dan brown.
We already have seen ai do racist and dei hallucinations, so its not far off to think it could start entrapoing or black mailing people once it achieves super intelligence
