Adversarial Manipulation of Reasoning Models using Internal Representations