[2312.02780] Scaling Laws for Adversarial Attacks on Language Model Activations