Currently, due to the significant negative impact of social bots, there has been widespread interest among researchers in automating the detection of social bots. And Graph Neural Network-based (GNN-based) detection methods have flourished, showing a very promising prospect in terms of detection performance. However, existing GNN-based social bot detection methods generally rely on densely annotated nodes in the context of social bot detection, leveraging them as training samples to guide the model training process, i.e., the detection social bot detection process. This demand for a large number of annotated nodes severely restricts the availability of GNN-based methods. To address this issue, we construct a GNN-based method that operates in a self-supervised pretraining-probing manner by stacking a graph masked autoencoder on top of a feature autoencoder (GMAE2). Benefiting from the pre-training of the encoder with self-supervised learning, the requirement of labeled nodes is significantly reduced. Through extensive experiments, we showcased that our GMAE2 is more suitable for social bot detection with an extremely low proportion of labeled nodes compared to existing methods. Our code is available at: https://github.com/CASIAhht/GMAE2-SBD.