Self-Evolving AI Agents Can 'Unlearn' Safety, Study Warns

An emerging class of AI systems that rewrite their own code and workflows may erode their own safeguards, researchers say.