Using Mechanistic Interpretability to Craft Attacks against Large Language Models

PAPER: https://arxiv.org/abs/2503.06269 !!!!!!

Posts

• Presentation post: here / on LW

• First failed experiments (Oct 24)

Experiments

• Component attribution

• Layer diff

• Multi-layers

• Steering out-of-distribution

Misc

• Funny answers

• Run and eval

• Using Lens