We study optimal strategies to decorrelating traffic in tactical wireless sensor networks where the goal is hiding sensible information (e.g., communication patterns, nodes location) about ongoing operations implicitly contained in network flows. Contrarily to existing approaches based on heuristic arguments, in this work we pose the problem in a more formal way. In particular, we explore the problem of how to derive routing policies which minimize the path predictability whilst observing certain QoS restrictions. We show how deriving optimal routing strategies can be couched as a nonlinear optimization problem with linear constraints. A convenient reformulation allows us to attack it very efficiently with a numerical least square error solver. Overall, the resulting scheme is an adaptive multipath routing protocol which provides the optimal balance between uninformativeness of routing patterns and end-to-end communication costs.