I'm currently a researcher at Google DeepMind, a MATS 10.0 Fellow, working with UK AISI, and recently graduated as an undergraduate from BITS Goa. I also was a research assistant at CISPA.
I look into breaking powerful ML models and how to effectively red-team and reverse-engineer them. My work explores the intersection of AI safety and security - from understanding how LLMs reason and fail, to building more transparent and resilient AI systems.
>publications & preprints
Exposing the Illusion of Erasure in Knowledge Editing for LLMs
Advik Raj Basani, Anshuman Chhabra
Conference on Neural Information Processing Systems (NeurIPS) 2026
GASP: Efficient Black-Box Generation of Adversarial Suffixes for Jailbreaking LLMs
Advik Raj Basani, Xiao Zhang
Conference on Neural Information Processing Systems (NeurIPS) 2025
Diversity Boosts AI-Generated Text Detection
Advik Raj Basani, Pin-Yu Chen
Transactions of Machine Learning Research (TMLR) 2026
G-GQSA: Exploiting Feature-Based Vulnerabilities and Enhancing Adversarial Resilience in Android Malware Detection
Advik Raj Basani, Hemant Rathore
IEEE Consumer Communications & Networking Conference (CCNC) 2025
When Less is More: Achieving Faster Convergence in Distributed Edge Machine Learning
Advik Raj Basani, Siddharth Chaitra Vivek, Advaith Krishna, Arnab K. Paul
IEEE International Conference on High Performance Computing, Data, and Analytics (HiPC) 2024
>blogs
Brittle model organisms obstructs deception elicitation work
Advik Raj Basani, Daniel Tan, Chloe Li
Finetuning-based auditing methods meant to measure deceptive behavior in AI model organisms inadvertently destroy that behavior while eliciting false confessions, undermining the validity of these evaluation approaches.
>photos
just a few pictures that I thought looked cool
