Research

My research is on agentic security in high-stakes domains, where LLM agents act with real authority and a single compromised action can cause irreversible harm.

A modern agent is more than its model: a harness wraps the model with prompts, tools, skills, and memory, and runs an agentic loop that plans, acts, and consumes untrusted inputs at every step. I study attacks across this stack — prompt injection, indirect prompt injection, prompt leaking, jailbreak, and memory poisoning — and defenses that keep agents secure once they are deployed in real systems.

My current work uses financial trading agents as a representative high-stakes setting; see SoK: Trading Agents or Market Crashers?.