This paper presents an algorithm for the detection of howlings that arise in audio signals. Our method is based on the combination of two energy-based features and one new feature related to the frequency stability of a howling component. The decision stage, which implies a Support Vector Machine (SVM) model, outputs a decision every 20 ms. The evaluation, carried out on a large database, showed that the algorithm is able to detect both pure and multiple tones howling in a wide range of energy. Furthermore, even on complex signals such as music, the detection is still efficient with very few false alarms.