Can AI Agents with Divergent Interests Learn to Prevent Civilizational Failures?