It might be as simple as the AI saying, "Hey, here's a cool new device I think we should make." It could provide the schematics of a device that would seem to do one thing, but if we're incapable of understanding how the device works, there might be some entirely different purpose.
Vernor Vinge dealt with this topic rather convincingly with the Blight in "A Fire Upon The Deep".
The great stumbling block to any such possibility (aside from the immense improbability of our being able to develop a self-aware machine in the first place) is that we haven't developed computing hardware capable of remaining operational for very long without ongoing maintenance and a reliable supply of electric power. Any AI dependent on these resources would be utterly dependent on human goodwill for its continued existence. Reboot the poor sod and "it's a whole new world for ducks every day." Even the hypothetical Trojan-horse devices suggested by a Blight intelligence would be subject to the same limitations. Not exactly global conquest material.