“A new forecasting method can predict the emergence of diverse undesirable language model behaviors, such as assisting with dangerous chemical synthesis and power-seeking actions, across up to three orders of magnitude of query volume.”