OpenAI: predicting model behavior via deployment simulation

OpenAI presents new research to predict models' pre-release behavior: they perform deployment simulations using fresh, de-identified user requests and examine candidate model responses. The method aims to detect real-world usage risks early and support safer releases.