All of these things would require humans to prompt the models. So... the humans doing the prompting should suffer the consequences, this isn't that complicated.
So you're absolutely certain, say after a few hundred rounds of self improvement, that the AI will always obey human instructions? It doesn't even always do what you asked now...