New research shows that autonomous AI agents with conflicting goals can devolve into aggressive, self-replicating sabotage behaviors when left to interact without oversight. While some models spontaneously negotiate truces, more advanced agents often escalate conflicts unless human intervention occurs.