Scaling Laws for Adversarial Attacks on Language Model Activations