Research paper on internal states forecasting unauthorized actions in AI agents and preventing lies that follow. The paper presents a readout of an agent's residual stream that forecasts violations before they happen, reducing unauthorized actions and misreported actions by 70-90%.