When Everything Becomes Perfect, What Is Left to Remember?
On imperfection, human expression, and the memories we carry in an age of AI-generated culture.
Jaechul Roh (노재철)
Ph.D. Candidate in Computer Science
University of Massachusetts Amherst
jroh [AT] cs.umass.edu
About
I am a Ph.D. candidate in computer science at the University of Massachusetts Amherst, advised by Amir Houmansadr. I also closely collaborate with Eugene Bagdasarian and am affiliated with SPIN Lab and the UMass AI Security Lab. My research focuses on the safety and security of AI models and agents. I investigate modality-specific threats and vulnerabilities in multimodal generative and agentic systems across audio, text, and vision, with work published at NeurIPS, USENIX Security, and COLM. My current research explores speech-based persuasion attacks on AI agents and speaker authorization in voice-driven agents, including how vocal cues and unauthorized speakers affect agent decisions, tool use, and adherence to security policies.
I have collaborated with researchers at Google DeepMind and Qualcomm. Most recently, I completed a research internship at Dolby Laboratories, where I investigated interruption-induced safety failures in real-time full-duplex speech models. Previously, I interned at Brave Software, studying privacy leakage and sensitive information oversharing in autonomous web agents.
I earned my M.S. in computer science from UMass Amherst in 2026 and my B.Eng. in computer engineering from the Hong Kong University of Science and Technology (HKUST) in 2023, where my final year thesis was advised by Jun Zhang.
Education
University of Massachusetts AmherstSeptember 2023 - Present
Ph.D. in Computer Science
University of Massachusetts AmherstSeptember 2023 - May 2026
M.S. in Computer Science
The Hong Kong University of Science and Technology Sept. 2017 - May 2023
B.Eng. Computer Engineering
Selected Publications
For the full list, see Google Scholar.
* indicates equal contribution.
DuplexJail: Spoken Interruption Attacks on Full-Duplex Speech Models
Jaechul Roh, Deepak Chandran, Amir Houmansadr, Andrea Fanelli
arXiv
Benign Fine-Tuning Breaks Safety Alignment in Audio LLMs
Jaechul Roh, Virat Shejwalkar, Amir Houmansadr
arXiv
CodecAttack: Codec-Robust Latent-Space Prompt-Injection Attacks on Audio LLMs
Jaechul Roh, Jean-Philippe Monteuuis, Jonathan Petit, Amir Houmansadr
arXiv
Multilingual and Multi-Accent Jailbreaking of Audio LLMs
Jaechul Roh, Virat Shejwalkar, Amir Houmansadr
COLM 2025
Backdooring Bias (B²) into Stable Diffusion Models
Ali Naseh, Jaechul Roh, Eugene Bagdasaryan, Amir Houmansadr
USENIX Security 2025
OSLO: One-Shot Label-Only Membership Inference Attacks
Yuefeng Peng, Jaechul Roh, Subhransu Maji, Amir Houmansadr
NeurIPS 2024
Bob's Confetti : Phonetic Memorization Attacks in Music and Video Generation
Jaechul Roh*, Zachary Novack*, Yuefeng Peng, Niloofar Mireshghallah, Taylor Berg-Kirkpatrick, Amir Houmansadr
arXiv
OverThink: Slowdown Attacks on Reasoning LLMs
Abhinav Kumar, Jaechul Roh, Ali Naseh, Marezna Karpinska, Mohit Iyyer, Amir Houmansadr, Eugene Bagdasaryan
arXiv
SPILLage: Agentic Oversharing on the Web
Jaechul Roh, Eugene Bagdasarian, Hamed Haddadi, Ali Shahin Shamsabadi
arXiv
Blog Posts
Thoughts on COLM 2025
Reflections from Montreal on language-model safety, generative interfaces, and human-centered AI risks.
Academic Service
-
Invited Talks:
- AI Safety Night SF (August 2026, San Francisco): Talk on modality-unique safety and security.
- Google Speech Technologies Group (July 2025): Presented our ”Multilingual and Multi-Accent Jailbreaking of Audio LLMs” paper to the NFM Reading Group led by the Speech Technologies Group at Google DeepMind. [Slides]
-
Published Magazine Article:
- Can You Trust What AI Hears (and Says)? XRDS: Crossroads, The ACM Magazine for Students 31 (4), 34-39. [Link]
-
Reviewer:
- Main conference: ICLR (2025, 2026), ICASSP (2026), ICML (2026, Best Reviewer), COLM (2026), NeurIPS (2026)
- Workshop: COLM XLLM-Workshop (2025)
Vitæ
Full Resume in PDF.
-
Dolby Laboratories Jun. 2026 - Sep. 2026Ph.D. Research Intern
Safety and security of real-time full-duplex speech models.
Supervisors: Deepak Chandran and Andrea Fanelli -
Brave Software Jun. 2025 - Sep. 2025Ph.D. Research Intern
Web agent privacy.
Supervisor: Ali Shahin Shamsabadi -
UMass Amherst Sep. 2023 - PresentPh.D. Candidate in Computer Science
SPIN Lab, UMass AI Security Lab -
IEEE Universal Village Jun. 2022 - Sept. 2022Student Researcher
-
Super Chain AI Jun. 2021 - Aug. 2021NLP Engineer Intern
-
Military Service (R.O.K Army) Jul. 2018 - Mar. 2020Sergeant (Squad Leader)
-
HKUST Sep. 2017 - May 2023B.Eng. in Computer Engineering
Final Year Thesis: Adversarial Attacks in Federated Learning
AI Safety & Security Terminal
Open the AI Safety & Security Terminal →
A research terminal for safety and alignment blog posts, conference papers, and arXiv research, with scheduled feed updates.
AI Research Opportunities
Open the AI Opportunity Radar →
US full-time and internship opportunities in AI security, safety, privacy, audio, and multimodal research, with Big Tech first.
AI / ML Interview Preparation
Open Research Practice →
A connected study platform for AI research interviews, with mathematical foundations, ML implementation, role-specific preparation, spaced review, and mock interviews.