This study examines the intersection of artificial intelligence and decentralized finance by proposing a network-based approach to modeling systemic financial risk in tokenized asset ecosystems. The conceptual foundation redefines risk as a relational property of graph structures rather than a scalar function, emphasizing the emergent behavior of decentralized financial protocols. The methodology integrates directional graph attention networks, specifically the DEDGAT architecture, trained on real and synthetic blockchain data to assess both inbound vulnerability and outbound contagion. Empirical findings demonstrate that the model significantly outperforms traditional benchmarks in identifying early-warning indicators, accurately detecting structurally critical nodes, and providing interpretable risk signals. The results confirm that directed graph embeddings enable more granular and adaptive risk stratification than symmetric or tabular models. The discussion of the findings situates this work within the latest literature on graph neural networks in finance, highlighting its contributions to regulatory visibility, DAO governance, and real-time monitoring. The conclusion affirms the effectiveness of DEDGAT as a scalable and policy-relevant tool for navigating the systemic fragilities of tokenized financial systems, while acknowledging the limitations imposed by synthetic validation data and current interpretability constraints.