Graph anomaly detection (GAD) aims to identify nodes that behave differently from the rest of the graph. It is an important task for many applications, from fraud detection to network security, and it becomes particularly challenging when only a small number of anomalous nodes are labeled - a common situation in practice, where anomalies are rare by nature and expensive to annotate. While unsupervised methods can achieve good performance without labeled data, they are unable to exploit that limited supervision that is often available in real-world scenarios. In this work, we study GAD under very few labels and propose a training augmentation framework based on Subgraph Forest Fire Sampling (SFFS).