Lessons Learned on Language Model Safety and Misuse

The deployment of powerful AI systems has enriched our understanding of safety and misuse far more than would have been possible through research alone. Notably: * API-based language model misuse often comes in different forms than we feared most. * We have identified limitations in existing language model evaluations that we are