Active Attention Probing

A new arXiv paper by researchers auditing internal safety scores finds that these scores anti-rank successful jailbreaks, with harmful generation rising from 0.05 to 0.27 on Llama while harmful intent…