Data center-level AI “kill switches” are more feasible than agent-level ones, but they need early detection and global regulatory alignment to work.