Automated vulnerability discovery in distributed systems

In this paper we present a technique for automatically assessing the amount of damage a small number of participant nodes can inflict on the overall performance of a large distributed system. We propose a feedback-driven tool that synthesizes malicious nodes in distributed systems, aiming to maximize the performance impact on the overall behavior of the distributed system. Our approach focuses on the interface of interaction between correct and faulty nodes, clearly differentiating the two categories. We build and evaluate a prototype of our approach and show that it is able to discover vulnerabilities in real systems, such as PBFT, a Byzantine Fault Tolerant system. We describe a scenario generated by our tool, where even a single malicious client can bring a BFT system of over 250 nodes down to zero throughput.