The `GLASS framework proposes a robust cross-domain transfer method based on graph language alignment for graph-level anomaly detection. This framework integrates structure-aware graph encoders with instruction-aware text embeddings into a representation space constructed from multi-slice soft cosine objectives, and serializes local, global, and semantic attribute sequences into compact graph descriptors (GraphDP). The model utilizes Matroshka representations of slices to enforce multi-scale consistency to capture anomalies at different granularities, and performs density estimation on the alignment hypersphere using spherical multi-modal scoring (SMS). Experiments show that GLASS achieves better average AUROC rankings than recent advanced baselines in twelve benchmark tests and three meta-domains, and demonstrates effective cross-domain transfer detection capabilities.