An outlier, known as an error state, can bring valuable cognitive analytic results in many industrial applications. Aiming at detecting outliers as soon as they appear in data streams that continuously arrive from data sources, this paper presents an adaptive-kernel-based incremental scheme. Specifically, the Gaussian kernel function with an adaptive kernel width is employed to ensure smoothness in local measures and to improve discriminability between objects. The dynamical Gaussian kernel density is presented to describe the gradual process of changing density. When new data arrives, the method updates the relevant density measures of the affected objects to achieve outlier computation of the arrived object, which can significantly reduce the computational burden. Experiments are performed on five commonly used datasets, and experimental results illustrate that the proposed method is more effective and robust for incremental outlier mining automatically.