Andis Draguns

Andis Draguns

Redteaming AI security research agendas

Selected Work

Unelicitable Backdoors via Cryptographic Transformer Circuits

NeurIPS 2024

Andis Draguns, Andrew Gritsevskiy, Sumeet Ramesh Motwani, Christian Schroeder de Witt

Demonstrates how backdoors can be seamlessly integrated into transformer models, questioning pre-deployment detection strategies.

Paper

Limitations of Agents Simulated by Predictive Models

ICLR 2024 Workshop

Raymond Douglas, Jacek Karwowski, Chan Bae, Andis Draguns, Victoria Krakovna

Outlines structural reasons why predictive models can fail when turned into agents, including auto-suggestive delusions and predictor-policy incoherence.

Paper

Residual Shuffle-Exchange Networks

AAAI 2021

Andis Draguns, Emīls Ozoliņš, Agris Šostaks, Matīss Apinis, Kārlis Freivalds

A lightweight network for processing long sequences that achieved state-of-the-art performance on music transcription.

Paper Code

Links