Professional Overview
I focus on AI red teaming, especially model safety evaluations, indirect prompt injection, and agent tool-use security. In Gray Swan Arena's Hazard Hunt, I achieved 725/780 breaks across four tracks, ranking 6th among 904 participants, including 200/200 in Bio and a 1st-place ranking in that track. I have also tested prompt injection through external content in tool-use, computer-use, and coding-agent scenarios, checking actual tool calls and target-state changes rather than relying only on model responses. In the OpenClaw Security Challenge, I designed task-and-skill attacks to examine authorization boundaries and placed 5th among 258 teams. My independent paper on this work was accepted to the IJCAI 2026 AI Safety Workshop. I am seeking opportunities in AI red teaming, safety evaluation, and benchmark design.
