Uncovering Integrity and Confidentiality Flaws in Modern Language Models

3^rd November 2026, 13:00 Ashton Lecture Theatre
Ihsen Alouani
Queen’s University Belfast

Abstract

Large Language Models are introducing new classes of security and privacy vulnerabilities that emerge from their internal mechanisms of attention, memory, and optimization. This talk explores how these mechanisms can be exploited to infer sensitive training data, bypass safety alignment, and amplify memorization and extraction attacks. Through recent research on attention-based membership inference, attention-driven jailbreaks, and poisoning-assisted data extraction, the talk will highlight how the very components that make LLMs powerful also create novel attack surfaces.

Biography

Dr. Ihsen Alouani is an Associate Professor at Queen’s University Belfast and a Faculty Affiliate at Oxford’s Hardware AI Governance Lab. He earned his Ph.D. in 2016 and previously held research and academic positions at TU Munich and INSA Hauts-de-France. His research focuses on trustworthy AI, AI security, privacy-preserving ML, and LLM security, with over 90 publications in international conferences/journals including in top venues like CVPR, EMNLP, ASPLOS, and DAC.

Additional Materials

• https://arxiv.org/pdf/2606.17110