AgentLens: Interpretable Safety Steering via Mechanistic Subspaces for Multi-Turn Coding Agent

AgentLens is a newly proposed white-box defense framework designed to enhance the safety of multi-turn coding agents using large language models (LLMs). It operates by detecting harmful execution states through step-level hidden representations and intervening in a 10-dimensional subspace within a single layer, marking a shift from traditional external guardrails. The introduction of the Mechanistic Agent Safety (MAS) benchmark, which includes annotated multi-turn execution trajectories from models like LLaMA-3.1-8B and Qwen-2.5-7B, demonstrates AgentLens's strong safety detection capabilities and its potential for risk anticipation, offering practitioners a novel approach to improve the safety of LLM applications in dynamic environments.