Yao Huang

The paper introduces TraceSafe-Bench, a comprehensive benchmark, and finds that securing LLM agents requires jointly optimizing for structural reasoning and safety alignment to mitigate risks during multi-step tool-use.

This paper analyzes and proposes four novel attack methods—based on model parameters and model inversion—to demonstrate that existing machine unlearning techniques can inadvertently leak the categories of the forgotten data.

MESA is a targeted alignment framework that decentralizes safety responsibilities across multiple experts in Mixture-of-Experts (MoE) LLMs using Optimal Transport theory, thereby improving safety robustness without sacrificing utility.