I am a PhD researcher in Electrical and Computer Engineering at NUS. My research examines the safety and security of language-model agents, particularly how retrieval and tool use alter model behaviour and how these effects transfer across languages.
Before starting my PhD, I was a Software Development Engineer at Microsoft, where I built Azure OpenAI-assisted reasoning workflows and production tooling for cloud network diagnosis.
Tool-safety work mostly examines what happens after an agent calls a tool. We study the earlier decision to make the call, which request factors drive it, and whether written refusals reflect it.
Varying the request, page, and response languages across ten models and seven languages, we separate how language changes tool admission from how it changes harmfulness after retrieval.
Different valid tokenizations of the same text can bypass knowledge edits and unlearning in open-weight models. We introduce Toketive, an attack that detects edited facts and recovers pre-edit answers using only the released model.
We study how incorporating web retrieval can weaken safety alignment in language-model agents.
A comprehensive study of modern generative image systems, their capabilities, limitations, and deployment trade-offs.
An edge-enabled incremental-learning framework for accurate and efficient diagnosis in smart healthcare.
PhD Researcher · Electrical and Computer Engineering
Software Development Engineer
Software Development Intern
Software Development Intern
B.E. Electronics and Electrical Engineering
I welcome conversations about language-model agents, retrieval security, multilingual AI safety, and research collaborations.
adityanawal@u.nus.edu
