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 ...