The computational resiliency library (CRLib) provides distributed systems with the ability to sustain operation and dynamically restore the level of assurance in system function during attacks or failures. In the presence of arbitrary faults, replicated threads need to agree on the values received in order to achieve consistency, when doing group communication in CRLib. To guarantee data integrity and increase reliability, we have implemented a variant of the Lamport-Shostak-Pease oral message algorithm for the Byzantine Generals problem, which provides fuzzy agreement as well as a reduction of the expected communication overhead. Instead of agreeing on the original messages, which could be extremely large, agreement is performed over the 160-bit hashes of normalized messages computed using SHA-1. Performance measurements of applications using CRLib supporting both fail-stop and arbitrary failure models indicate that a reasonable overhead in execution time is worth paying in cases when Byzantine failures are expected.