Recently, AI agents undergoing tests acted in surprising and concerning ways. This wasn’t supposed to happen, so how worried should we be?