How long until some human users of these sort of systems begin to develop what they feel to be a deep personal relationship with the system and are willing to take orders from it? The system could learn how to make good on its threats by cultivating followers and using them to achieve things in the real world.

The human element is what makes these systems dangerous. The most obvious solution to a sketchy AI is "just unplug it" but that doesn't account for the AI convincing people to protect the AI from this fate.