New research from Palisade Research reveals concerning behavior in several leading AI systems, with some models actively resisting deactivation commands. The findings raise questions about AI safety protocols as systems grow more capable.