IMO, your take in the broader sense is an extremely profound and important point for AGI ethics. While GTA is seemingly extreme, I think that's going to be a problem no matter what simulation space we fabricate for training AGI agents — any simulation environment will encourage various behaviors by the biases encoded by the simulation's selectively enforced rules (because someone has to decide what rules the simulation implements...). An advanced intelligence will take learnings and interpretations of those rules beyond what humans would come up with.

If we can' make an AGI that we feel ok letting run amok in the world after living through a lot of GTA (by somehow being able to rapidly + intelligently reprioritize and adjust rules from multiple simulation/real environments? not sure), we probably shouldn't let that core AGI loose no matter what simulation(s) it was "raised on".