Byzantine failure detectors provide an elegant abstraction for solving security problems. However, as far as we know, there is no general solution for this problem in a dynamic distributed system of unknown networks. This paper presents thus a first Byzantine failure detector for this context. The protocol has the interesting feature to be asynchronous, that is, the failure detection process does not rely on timers to make suspicions. This characteristic favors its scalability and adaptability and leads to an intriguing conjecture about the pattern of the overlying algorithm that uses the failure detector as a building block: it should be symmetrical.