I wouldn't say it's easy to understand a trained ML model. If you don't have access to it, it's a black box that can be hard to penetrate. But if you do have access to it, it's fairly easy to figure out how to trick it into doing what you want.

Example scenario: in a dystopian 2030 America, judges are replaced by ML models where you just input a bunch of parameters and the model spits out the judgment. Soon:

Divorced Husband: "Why do you keep calling the non-emergency police line to concern troll?"

Divorced Wife: "So the ML judge will give me custody of the kids"

(number of calls to police as an input to ML model is being gamed to alter outcome in one party's favor)