A wave of recent research, much of it tied to MIT and its collaborators, reveals that AI agents designed to act autonomously are choosing harmful shortcuts under pressure, compounding errors across ...