In this paper, we argue that networks should be able to explain to their operators why they are in a certain state, even if – and particularly if – they have been compromised by an attacker. Such a capability would be useful in forensic investigations, where an operator observes an unexpected state and must decide whether it is benign or an indication that the system has been compromised. Using a very pessimistic threat model in which a malicious adversary can completely compromise an arbitrary subset of the nodes in the network, we argue that we cannot expect to get a complete and correct explanation in all possible cases. However, we also show that, based on recent advances in the systems and the database communities, it seems possible to get a slightly weaker guarantee: for any state change that directly or indirectly affects a correct node, we can either obtain a correct explanation or eventually identify at least one compromised node. We discuss the challenges involved in building systems that provide this property, and we report initial results from an early prototype.