Why can’t we just keep rogue AIs off the internet?

Researchers test AI agents in controlled environments to predict and mitigate their potential risks, but some argue that isolating them from the internet could hinder their development and understanding of their behavior.