Agent MechSuits: Mechanistic Subspace Safety Steering for Multi-Turn CLI Agents

Weidi Luo, Qiming Zhang, Yihao Quan, Mingyu Jin, Jie Cai, Chaowei Xiao, Jingcheng Niu and Zhen Xiang.
NeurIPS 2026

arXiv

TL;DR

Agent MechSuits is a white-box defense that detects and mitigates unsafe behaviour of CLI agents at runtime, directly in the model’s representation space. Rather than relying on external guardrails, it intervenes within a low-dimensional (10-d) safety subspace during multi-turn execution. We also introduce the Mechanistic Agent Safety benchmark: annotated multi-turn trajectories across 194 tasks and three model families. MechSuits achieves strong detection performance and substantially reduces harmful agent actions.

How to Cite

@inproceedings{luo2026mechsuits,
  title     = {Agent MechSuits: Mechanistic Subspace Safety Steering for Multi-Turn {CLI} Agents},
  author    = {Luo, Weidi and Zhang, Qiming and Quan, Yihao and Jin, Mingyu and Cai, Jie and Xiao, Chaowei and Niu, Jingcheng and Xiang, Zhen},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026},
}