Evaluating Language Models for Harmful Manipulation