This paper addresses a reformed observer-based fault tolerant control problem for uncertain Markov jump delayed systems (MJDSs) with unknown state delay, actuator failures (AFs) and uncertain transition rates (TRs) under a novel sliding mode framework. Meanwhile, an neural network (NN) based adaptive output feedback controller is proposed to mitigate potential AFs and unpredictable attacks. The key novelty of this study is that the devised control method could guarantee the reachability of sliding surface in limited time and further stabilize the MJDSs stochastically in spite of possible AFs, unknowable delay, uncertain TRs and uncertain nonlinearity simultaneously, which provides an alternative way to state-estimator-based controller design for the MJDSs, especially when the accurate knowledge of state delay cannot be obtained. At last, an RLC circuit model is introduced to verify the effectiveness of the theoretical result.