Does Self-Evaluation Enable Wireheading in Language Models?