01 — RESEARCH PHILOSOPHY
Breaking Systems to Build Better Ones
Red teaming is not just about attack—it's about defense.
True robustness comes from first principles, not band-aid fixes.
M2S: Multi-turn to Single-turn jailbreak in Red Teaming for LLMs
96% ASR, 60%+ token reduction. Three templates: Hyphenize, Numberize, Pythonize.
Human's Last Exam (HLE)
Challenging benchmark dataset for evaluating frontier LLM capabilities.
X-Teaming Evolutionary M2S: Automated Discovery of Multi-turn to Single-turn Jailbreak Templates
Automated M2S template discovery with LLM-guided evolution. 44.8% success on GPT-4.1.
ObjexMT: Objective Extraction and Metacognitive Calibration for LLM-as-a-Judge under Multi-Turn Jailbreaks
LLM-as-a-Judge benchmark for hidden objective extraction. 47-61% accuracy across models.
03 — WORK EXPERIENCE
Building at the Frontier
AIM Intelligence
- ▸First author, ACL 2025 Main Conference - M2S framework
- ▸Led KT Internal LLM Evaluation benchmarking initiative
- ▸Qualification Round Lead, KISA AI Hacking Defense Competition 2025
- ▸Red Teaming dataset initiative for major Korean telecom operator
Coupang Eats
- ▸Built Python automation tools boosting productivity 10-300%
- ▸Developed menu-matching ML model (93-94% accuracy)
- ▸Automated Presto ↔ Google Sheets pipelines
EDUCATION
University of Seoul
MathematicsMar 2018 - Present
COMPETITIONS
LLM Jailbreak Challenge
1st place (solo)AIM Intelligence
Independently developed a jailbreak solution that bypassed the model’s tightly engineered internal guidelines and successfully extracted protected internal keys from the LLM.
Adversarial Attack on Vision Models
2nd placeAIM Intelligence
Designed minimal image perturbations that caused the model to predict “keep driving” despite a pedestrian ahead, successfully bypassing the target vision model’s safety defenses and earning 2nd place.
03 — TECHNICAL EXPERTISE