I am currently working in Meta Superintelligence Labs as a Research Scientist. I work on safety, robustness and alignment. I completed my PhD at Carnegie Mellon University in 2023, advised by Anupam Datta and Matt Fredrikson. I have been doing research in the industry since my graduation. Before Meta, I also worked at Scale AI and the Center for AI Safety.
Selected Work
A comprehensive report for Muse Spark on preparedness, adversarial robustness, and model behaviors.
A monitor red-teaming workflow for detecting covert misbehavior in autonomous LLM agents, including weak-to-strong monitoring gains.
A red-team workflow that turns capable LLMs into attackers for finding jailbreaks against other models.
A universal adversarial suffix attack that transfers across aligned language models and exposes failure modes in public LLM interfaces.
A direct PAC-Bayesian training objective that tightens robust generalization bounds while improving adversarial robustness.
A training objective and architecture adaptation for certifiably robust models using efficient global Lipschitz bounds.