An algorithm for robust fitting of AR models is given, based on a linear regression idea. The new method appears to outperform the Yule-Walker estimator in a setting of data contaminated with outliers.