The paper presents a method for automated vehicles, with which safety critical maneuvers can be detected. The proposed method is based on state observation of lateral dynamics, which functionality in a novel observer structure is achieved. The observer contains robust model-based and reinforcement learning-based (RL) methods, which in a performance guaranteed structure are combined. Simulations in the paper illustrate that losses of lateral stability and tracking performance the proposed safety evaluation method can be detected, which is a novel contribution compared to the existing stability index based methods.