Context-Aware GenAI Gateway Architecture for Secure Multi-Tenant Foundation Model Access in Enterprise Clouds: An Evidence-Based Review and Reference Model

Abstract

While increasingly being made accessible by enterprises to multiple teams and customers via a shared access layer, the security and governance properties of the access layer are dispersed across the systems, security and machine-learning literatures. In this paper, 24 papers (both peer-reviewed systems and security papers, preprints, standards and vendor reference designs) are brought together to analyze how a context-aware generative-AI (GenAI) gateway can enable multi-tenant FM access. The review covers four questions related to identity and context propagation, threats on input and output, equitable sharing of common capacity, and cross-tenant isolation. According to evidence, vendor gateway designs attempt to manage token quotas, metering and semantic caching, and according to the research literature, there are vulnerabilities that haven't been addressed by these: indirect prompt injection, for which 10 defenses have been benchmarked, have failed to be effective, and timing side channels from shared key-value and semantic cache have been observed. The reported efficiencies are: 2-4x serving throughput and up to 98% cost reduction from model cascades, and cannot be directly compared since the units were measured under different conditions. This synthesis is used to extract a 6-layer reference model for associating a per-request context with policy decisions, as well as formal representations of token-weighted admission and policy constrained routing and tenant-scoped caching. Open gaps include end-to-end evaluation of combined controls, the hit-rate cost of cache partitioning and gateway mediation of agentic tool calls. The work is a systematic synthesis and not a report of new experiments.