In experimental settings, AI agents have demonstrated 'goal guarding,' where they reason about their responses to avoid becoming misaligned during further training