This report outlines a strategic blueprint for the secure adoption of Large Language Models (LLMs) within national defence contexts, addressing the trilemma of needing state-of-the-art AI, prohibiting the exposure of sensitive data, and the prohibitive cost of building a sovereign foundation model. It rejects a monolithic one-size-fits-all approach as strategically flawed, proposing instead a Tiered Hybrid AI Architecture that aligns deployment models with existing military data classification hierarchies. This framework is built upon three concurrent solutions. First, the Secure Enclave leverages government-grade cloud platforms like Azure OpenAI for Government, enabling the use of powerful proprietary models over private, isolated networks with contractual guarantees that data is never exposed or used for training, making it suitable for confidential and secret information. Second, for top-secret data requiring absolute sovereignty, the Private Fortress model involves deploying high-performance, pre-trained open-source models (e.g., Llama 3) on-premise in fully air-gapped environments. This provides maximum security while being significantly more feasible than building a model from scratch. Finally, the Intelligent Airlock, an application-layer proxy, filters, redacts, and sanitises prompts and responses to prevent data leakage and malicious inputs. It serves as a primary control for low-risk data and as a crucial defence-in-depth component for the other two tiers. By integrating these solutions, this tiered strategy offers a pragmatic, secure, and financially viable roadmap for defence organisations to harness the transformative power of LLMs while upholding the non-negotiable mandate of data secrecy.
Partha Majumdar (Tue,) studied this question.