Anyone else realize giving an AI more access can make things harder to control?

When I first started using agents I thought the answer was simple give them more tools and they can do more useful work. But the more access they get the more I start thinking about permissions costs data and what happens when something goes wrong. Having more capability also means having more things to keep an eye on. Where do you draw the line between making an agent powerful enough to be useful and giving it too much freedom?