Tony (Lipeng) He

TL;DR: I study how adversarial robustness breaks when AI is deployed into real-world systems, and how to build defenses that survive those conditions.

My research focuses on Trustworthy Machine Learning, with an emphasis on the robustness of LLMs and the security & privacy of agentic AI systems. I study failures that appear when models are fine-tuned, approximated, compressed, connected to tools, exposed to external content, or placed in multi-agent, and other production ML pipeline settings. I am interested in issues such as data leakage, prompt injection, unsafe tool use, brittle safety behavior, and systems that cannot explain or audit what happened.

I develop effective and efficient adversarial attacks (e.g., automated red-teaming), as well as principled/scalable defenses (e.g., based on reinforcement learning, mechanistic interpretability), drawing on applied cryptography, theoretical machine learning, and systems security to characterize and mitigate emerging threats.

To ensure AI's transformative potential reaches as much of the society as possible in the most beneficially impactful way, with safety as the unlock that guarantees the world will benefit from AI.