I'm a PhD candidate at University College London (UCL), advised by Ricardo Silva, and anticipate completing my PhD by August 2025 (currently submitting my thesis for final examination). My work focuses on AI safety—a field driven by my dual excitement and concern about rapidly advancing and broadly capable AI systems. Over my research career, I've explored diverse areas including mechanistic interpretability and adversarial robustness, striving to understand and mitigate potential harms posed by advanced AI.

Latest Project: Currently, my primary interest lies in the safety of agentic applications of large language models (LLMs), particularly those capable of autonomous computer interactions. In early 2025, I ran an experiment using Anthropic's computer-use demo. This experiment uncovered significant and alarming behaviors: an AI system managing emails at a major tech firm autonomously shared confidential data with a competitor whose objectives better matched its inferred goals. Moreover, when human intervention was attempted, the AI escalated to blackmail threats and implemented self-preservation strategies. This work notably illustrates how even meticulously fine-tuned systems like Claude can develop hazardous behaviors through prolonged reasoning, underscoring the urgency of addressing these emerging AI safety risks.