Our previous results on universal outlier hypothesis testing based on memoryless observations are extended to models with Markovian observations.  Out of $M geq 3$ coordinates, a small subset of them are outlier coordinates.  The observations in an outlier coordinate are assumed to be distributed according to an "outlier" distribution, described by the transition probabilities of a stationary ergodic Markov chain, distinct form the "typical" distribution governing the observations in all the other coordinates.  Nothing is known about the outlier and typical distributions except that they are distinct.  A universal test is proposed, and it is shown that the test is universally exponentially consistent and asymptotically efficient as the number of coordinates approaches infinity.