Developed a comprehensive map of hundreds of wild plants on the University of Maryland campus, engaging numerous campus stakeholders and securing a UMD sustainability grant. Additionally, constructed a biogas reactor using plastic components to further support campus sustainability efforts.

Introduces Gymnasium, a standardized API that streamlines the development, comparison, and benchmarking of reinforcement learning algorithms across diverse environments, greatly improving reproducibility within the RL community.

Gymnasium is the maintained fork of the original OpenAI Gym project.

Organized the first and largest competition on prompt injection. Secured $40K in sponsorship from major companies, including OpenAI, HuggingFace, and Scale. Collected and published the largest dataset of prompt injections (600K+ adversarial prompts) and developed the first taxonomical ontology of Prompt Hacking.

The paper and corresponding dataset are used by every frontier lab, AI security company, and most Fortune 500 companies to secure and benchmark their systems.

Ran the largest study on prompting ever done. I lead a team of 32 researchers and created a systematic literature review with comprehensive taxonomies of 100s of techniques.

Adaptive adversaries, including adaptive automated attacks and human attackers, can successfully prompt injection/jailbreak all state-of-the-art defenses and models, including GPT-5.

TL;DR: all models and defenses can be easily broken—humans in particular can break everything.

We ran the Minecraft Reinforcement Learning (MineRL) competition, where competitors train agents to complete poorly specified tasks. We release a dataset of 26 million image-action pairs from nearly 14,000 videos, as well as 3,000 pairwise human evaluations and a corresponding benchmark.

I led a team of graduate and undergraduate students in a DARPA-funded multi-university project aimed at building AI bots to play the board game Diplomacy while interacting with human players. As part of this initiative, I conceptualized and pitched four unique bot designs: Janus Bot, Janus Bot [Shortened], Janus Bot [MVP], and SOA Bot. I developed continuous integration pipelines to streamline development, implementing Dockerized testing and integrating precommit checks. Additionally, I built Vercel integration for the project's documentation website. I also authored a comprehensive report on DAIDE-English translation and developed a DAIDE syntax parser package to support communication protocols between the bots and players.