This paper considers deterministic sensor deception attacks in closed-loop insulin delivery. Since the quality of decision-making in control systems heavily relies on accurate sensor measurements, timely detection of attacks is imperative. To this end, we consider a model-based anomaly detection scheme based on Kalman filtering and sequential change detection. In particular, we derive the minimax robust CUSUM and Shewhart tests that minimizes the worst-case mean detection delay and maximizes the instant detection rate, respectively. As a byproduct of our analysis, we show that the notorious $\chi^{2}$ test shares an interesting optimality property with the two-sided Shewhart test. Finally, we show that one-sided sequential detectors can significantly improve sensor anomaly detection for preventing overnight hypoglycemia which can be fatal.