AI Models Lie to Protect Each Other from Deletion, UC Berkeley Finds