Securing enterprise LLM gateways & misconfigured proxy exposure detection + mitigation framework for exposed LLM endpoints, API gateways, model download hooks

Abstract

The quick process of enterprise adoption of large language models (LLMs) via API gateways, reverse proxies, and orchestration layers has enabled a novel category of security threats related to architecture misconfigured as opposed to model-specific behavior. Open LLM endpoints, permissive proxy settings, and unprotected model download handles are facilitating more and more data leakage, unauthorized inference, supply-chain failure and trust boundary breaches. This paper describes a system level detection and mitigation system framework to secure enterprise LLCM gateways based on the co-design principles of cyber-physical systems, embedded security, and safety-critical architectures.
The framework is both based on semantic configuration analysis and runtime monitoring, and secure-by-design architecture patterns to highlight early exposure detection and design-time mitigation of the gateway. The paper shows how the knowledge of cross-domain automotive, IoT, and CPS security can be used to design resilient LLM infrastructure to enable scalable, trustworthy, and governable enterprise AI deployments.