International Revenue Sharing Fraud (IRSF) is the most persistent type of fraud in the telco industry. Hackers try to gain access to an operator’s network in order to make expensive unauthorized phone calls on behalf of someone else. This results in massive phone bills that victims have to pay while number owners earn the money. Current anti-fraud solutions enable the detection of IRSF afterwards by detecting deviations in the overall caller’s expenses and block phone devices to prevent attack escalation. These solutions suffer from two main drawbacks: (i) they act only when financial damage is done and (ii) they offer no protection against future attacks. In this paper, we demonstrate how unsupervised machine learning can be used to discover fraudulent calls at the moment of their establishment, thereby preventing IRSF from happening. Specifically, we investigate the use of Isolation Forests for the detection of frauds before calls are initiated and compare the results to an existing industrial post-mortem anti-fraud solution.