It may already be too late to control AI

The report’s most bracing shift from the year before comes through a simple pattern: capability gains keep widening the number of harm pathways, while real-world visibility into misuse grows much more slowly.