新智元 (微信公众号)T2·Agents
72 pts
A research team from the University of Georgia, Rutgers, USC and others proposes Agent MechSuits, which reads internal safety signals before each action of a CLI agent and applies sparse intervention on just 10 hidden representation dimensions in a single layer while keeping the base model frozen. Risk detection on the current step reaches 97.32% average accuracy across three tested models, and the paper is accepted to NeurIPS 2026.