Investigating Sensitive Directions in GPT-2: An Improved Baseline and Comparative Analysis of SAEs