The neighbor-based detection technique explores the principle that sensor nodes situated spatially close to each other tend to have a similar behavior. A node is considered malicious if its behavior significantly differs from its neighbors. This detection technique is localized, unsupervised and adapts to changing network dynamics. Although the technique is promising, it has not been deeply researched in the context of wireless sensor networks yet. In this paper, we present symptoms which can be used in the neighbor-based technique for detection of selective forwarding, jamming and hello flood attacks. We implemented an intrusion detection system which employs the neighbor-based detection technique. The system was designed for and works on the TinyOS operating system running the Collection Tree Protocol. We evaluated accuracy of the technique in the detection of selective forwarding, jamming and hello flood attacks. The results show that the neighbor-based detection technique is highly accurate, especially in the case when collaboration among neighboring nodes is used.