What if AI decides humans need to be terminated?

Inspired by:
https://www.reddit.com/r/ControlProblem/comments/53f4h5/is_it_possible_that_a_dangerousforhumanity/
Imagine a self-improving hyperintelligent AI that finds humans are a problem or an obstacle. How this can be stopped?

why are so people afraid of murderous AI? what can a self-improving hyperintelligent AI do if it's imprisoned inside an aging lith machine, a university mainframe or an old crackpot's chatbot?