Concrete Problems in AI Safety Part 2: Empowerment - eviltoast

Maybe AI systems would be safer if they avoid gaining too much control over their environment? How might that work?