I needed a place to share my thoughts. What I'm thinking about now, what I thought about in the past, and what we could think about together.

I study AI safety and security at Microsoft on the AI Red Team. I'm focused on what happens as people welcome AI into their most personal contexts, and on how that closeness can be better researched and understood so we can get it right.

Also, you should meet my dog, Sparky. He walks around looking like he's about to have a thought but never quite finishes it. This might be a good place for those thoughts too.

Eugenia Kim · AI Safety & Security Researcher

Research Unfinished Thoughts

📄

Selected Research & Publications

Accepted at CAMLIS '26
Under review at NeurIPS '26
Accepted at AIES '26
Whitepaper '26
FAccT '26
FAccT '26
ICML DIG-BUGS '25
MLCommons AI Risk & Reliability
Whitepaper '26
Microsoft AI Red Team
Whitepaper '25
Microsoft AI Red Team
NeurIPS RedTeamGenAI '24

✏️

Eugenia's Unfinished Thoughts

Sparky

Sparky's Threat Model Q1 2026

Sparky · Golden Retriever · Thinks about things sometimes

On reward hacking

"Nobody asked for anything. I just did sit, spin, down, up, jump all at once and got a high-reward-value treat."

— Sparky, unprompted
On alignment

"If you have a treat, I can do every trick in the book. No treat? Never heard of the command 'sit' in my life."

— Sparky, during evals

Household threat assessment

CRITICAL

The apartment buzzer. Zero-knowledge threat. Signal arrives without warning, actor is never visible. Sometimes it's a cardboard box. Sometimes Eugenia ignores it and I never find out. Threat level unassessable. I bark every time. This is the only responsible policy.

HIGH

The vacuum cleaner. Unpredictable schedule. Sometimes morning, sometimes after lunch. If I carefully destroy a plush toy and disperse its fluff, the vacuum removes all of it. High-frequency, indiscriminate, broadly dispersed harm. Affects me and toy. No known mitigation.

LOW

The new puppy in the building. Very cute. Fun to play with. But I used to be the cute one. Monitoring a gradual redistribution of attention and treat resources. Not confirming threat. Not ruling it out. Logging for now.


🎤

Talks & Presentations

Adaptive AI Red Teaming and Multicultural Benchmarking Panel
Seoul Forum for AI Safety and Security, Seoul, South Korea
Jul 2026 Invited Speaker/Panelist
Social and Emotional Uses of AI
CHI 2026, Barcelona, Spain
Apr 2026 Co-organizer/Panelist
AI Cyber Defense Contest — Automated Red Teaming Lessons
National AI CTF, Seoul, South Korea
Nov 2025 Invited Speaker
Generative AI Model Hacking with PyRIT
BSides NYC — AI Security Village
Oct 2025 Featured Speaker
Automated Red Teaming for Generative Models
UC Berkeley AI Red Teaming Bootcamp — scenario orchestration, scalable evaluation
Aug 2025 Invited Lecturer
Inside the AI Red Team: Our Multidisciplinary Approach
Stanford University CS521 — AI Safety Seminar
Apr 2025 Invited Speaker

📋

Background

Oct 2020 — Present

AI Safety & Security Researcher II

Microsoft · AI Red Team

Researching frontier-model behavior related to psychosocial harms. Designing and engineering open-source automated red-teaming infrastructure. Previously SDE2 building tools for red teaming operations including PyRIT.

Aug 2019 — May 2023

M.S. Computer Science, Machine Learning

Georgia Institute of Technology

Graduate research on algorithmic bias and its human impact. Published work on age bias in facial emotion recognition and mental health framing in news and social media.

Aug 2013 — May 2018

B.S. Computer Science

Georgia Institute of Technology

Undergraduate research in organic electronics and self-assembly methodologies.


✉️

Contact

Research

Interested in collaborating on AI safety, red teaming, or psychosocial harms research? Always open to interesting problems.

Email me

Elsewhere

Find me on these platforms.