Geoffrey Irving, Chief Scientist at the UK AI Security Institute, explains why our theoretical understanding of machine learning remains fragile even as models surpass experts on critical security tasks. He details AISI’s work on frontier model evaluations...