It would be easy for a superintelligent AI to kill off humans, and we'd have no idea it was happening until it did. Most likely it would use human gullibility to get us to act in its interests until there was suitable infrastructure in place that we were no longer necessary. Social media has demonstrated how easily people are fooled and the tools at the disposal of an AGI would be far more powerful. There are some excellent explanations of why this risk is real, written by people who've been working in the field for years.