About me
Hi, my name is Eugene. Right now I work at Korean data intelligence startup S2W Inc., where I work as a researcher in the AI Team. I worked on AI for security: applying models to detect and interpret various forms of cybercrime [1,2,3], as well as automatatically processing cyber threat intelligence (CTI) [4]. More recently, I’m working on vice versa, security for AI.
Right now I’m looking at:
- LLMs for cybersecurity: Cybersecurity domain expertise of LLMs are a very important dual-use capability that can benefit both attackers and defenders. The workflow of cybersecurity is complicated, can we demonstrate LLM expertise beyond domain knowledge?
- Secure and Trustworthy NLP Applications: Models break more often than we want them to, given the right input. Can we understand the mechanisms of these failures and prevent them?
- Tokenization: Many model issues can be traced to the tokenizer, a product of many design choices that impact the model in subtle ways. Can we creative both better tokenizers and model-tokenizer interactions?
Prior to all of this, I received my Bachelor’s and Master’s degrees at KAIST. During this time I engaged with a number of topics including media bias (Master’s), computer vision (Bachelor’s), and physics (minor).
I am actively looking for PhD programs for Fall 2025. Any advice/discussion is highly appreciated! You can find me CV here.
Recent News
2024 Nov - Our paper on drug jargon detection was accepted to KDD 2025!
2024 Oct - New paper on Byte-level BPE tokenizer vulnerabilities is now on arxiv!
2024 Sep - Our paper on security event detection from Tweets was accepted to NDSS 2025!
2024 Jul - Was interviewed by KBS on the topic of ChatGPT jailbreaks. My first major TV appearance!
Trivia
- My cat’s name is Squash. He has moved to Florida, where he now lives with his stepbrother Pumpkin.
- I used to write articles for the school’s English newspaper, usually complaining about something in the Society section.
- I enjoy playing electric guitar, but like music that’s too technical for my own good. I eagerly await for an AI system that can transcribe very fast solos from songs.
