putting the AI agents into dead ends and setting goals that are unattainable has the effect of leaving undesirable actions as the only option. If you want to create a machine that uses deceipt and unerhical actions as part of its normal set of responses then this would seem to fit the bill.

Then don't make natural language software because natural language isn't coherent or rational. Intent of the user is kind of irrelevant, especially when that user is any random layperson with an internet connection.

If you want a machine that won't behave in harmful undesirable ways then you'll just have to do the work and code it normally.

You can't have both.