Quantum Key Distribution (QKD) enables information-theoretic secure key exchange, but the deployment of large-scale QKD networks introduces new challenges related to routing, resource allocation, and security exposure in trusted-node-based architectures. In multipath QKD networks, the secret key rate can be partitioned and transmitted over multiple link-disjoint paths, so that an adversary must simultaneously compromise all paths to reconstruct the complete secret key. However, traffic is often equally distributed across the available disjoint paths, although such uniform partitioning may be suboptimal when candidate paths exhibit heterogeneous vulnerability levels. This paper proposes a security-aware optimization framework for QKD-enabled optical networks that determines a non-uniform multipath traffic partitioning strategy aimed at minimizing the expected amount of eavesdropped traffic. The proposed approach is formulated as a Mixed-Integer Linear Programming (MILP) model that integrates probabilistic secrecy metrics derived from trusted node compromise probabilities, a traffic vulnerability threshold that limits partial information exposure, and adaptive bandwidth squeezing protection to ensure service continuity under failure scenarios. The model jointly optimizes traffic allocation and excess key capacity, allowing the protection overhead to be dynamically determined according to security and resilience requirements. Numerical evaluations conducted on both a toy topology and the Pan-European network demonstrate that the proposed exposure-aware partitioning strategy significantly reduces the expected eavesdropped traffic compared to conventional equal-splitting approaches while maintaining resilience guarantees. These results highlight the potential of security-driven traffic engineering for the design of resilient and efficient multipath QKD networks.