Dodis et al. proposed an improved version of the fuzzy vault scheme, one of the most popular primitives used in biometric cryptosystems, requiring less storage and leaking less information. Recently, Blanton and Aliasgari have shown that the relation of two improved fuzzy vault records of the same individual may be determined by solving a system of non-linear equations. However, they conjectured that this is feasible for small parameters only. In this paper, we present a new attack against the improved fuzzy vault scheme based on the extended Euclidean algorithm that determines if two records are related and recovers the elements by which the protected features, e.g., the biometric templates, differ. Our theoretical and empirical analysis demonstrates that the attack is very effective and efficient for practical parameters. Furthermore, we show how this attack can be extended to fully recover both feature sets from related vault records much more efficiently than possible by attacking each record individually. We complement this work by deriving lower bounds for record multiplicity attacks and use these to show that our attack is asymptotically optimal in an information theoretic sense. Finally, we propose remedies to harden the scheme against record multiplicity attacks.