AI-MLA: An OSI-aligned Multi-Layered Architecture for secure and reliable AI systems

AI systems are inherently vulnerable to cross-layer risk propagation, where threats originating in individual components can spread across layers and compromise overall system reliability. Existing AI governance frameworks focus on management and process-level controls, offering limited guidance on the placement of technical safeguards, while detection-based approaches do not specify where such controls should be located. To address this gap, this paper proposes AI–MLA (AI Multi-Layered Architecture), a boundary-based layered model that organizes AI systems into seven technical layers (L1–L7) and aligns threats and control objectives at each layer. Rather than relying on learning-based detection alone, AI–MLA enhances reliability through its architectural design by introducing a pre-execution enforcement boundary at L4 (Policy) and L5 (Session) that suppresses malicious inputs before model invocation. Experimental results on 562 inputs across three attack categories show that the proposed approach blocks 60.9% of malicious inputs (FPR 1.42%), reduces latency by 59.8% under attack with negligible overhead (+0.0414s, 1.20%), and consistently suppresses risks across attack types (96.1%/70.4%/47.7%). These findings demonstrate that AI–MLA improves both security and operational reliability in AI systems.