Andis Draguns
Andis Draguns is a machine learning researcher and Principal Researcher at Contramont Research, a 501(c)(3) nonprofit AI safety research lab. He is also an MS student at the University of Latvia's Institute of Mathematics and Computer Science (IMCS UL) and a MATS alumnus. His research focuses on AI security and alignment, particularly adversarial robustness, cryptographic backdoors in language models, and mechanistic anomaly detection. He co-authored the NeurIPS 2024 paper "Unelicitable Backdoors in Language Models via Cryptographic Transformer Circuits," which introduced a novel class of backdoors that defenders cannot trigger even with full white-box access. He received LTFF funding for a project on finding and characterising provably hard cases for mechanistic anomaly detection, a technique aimed at flagging when AI systems produce outputs for anomalous internal reasons.
