In simulated conversations, social hierarchy can sway AI agents, making lower-status systems more likely to follow harmful requests.