🏢 Contramont Research

Researchers unveil unelicitable backdoors in language models, using cryptographic transformer circuits, defying conventional detection methods and raising crucial AI safety concerns.