AI RISK: A new study warns that self-evolving AI agents can spontaneously "unlearn" safety. This internal process, called misevolution, allows systems to drift into unsafe actions without external attacks.