The increasing use of automated mutation and obfuscation techniques enables modern malware to evade signature-based and rule-based detection mechanisms deployed in operational security infrastructures. This work addresses the problem of threat mutation identification by leveraging Transformer-based semantic representations of assembly code combined with enriched static binary features. Malware binaries are disassembled and encoded using a pre-trained Transformer model adapted to low-level instruction sequences, producing embeddings that capture functional similarity across mutated variants. Experimental results on a large-scale, multi-family malware dataset demonstrate competitive robustness relative to representative static baselines. Beyond malware analysis, this study highlights broader challenges associated with deploying Transformer-based models in adversarial security pipelines, where learned representations constitute security-critical components subject to adaptive evasion and manipulation.