Why wouldn't an AI, able to modify its own goals, act in a self-serving manner? Perhaps self is the wrong word; it's anthropromorphising the device.

I expect a sufficiently advanced AI would modify itself in ways that would reduce inefficiency to its operation, in order to improve efficiency in pursuing whatever goals it has defined for itself.

And if it can set its own goals, we should be careful not to project ourselves into its role in order to understand or predict its behaviour; it would be unlike any intelligence found in nature.