CRAGI - Cybersecure RAG Implementations

This proposal targets focus areas 1, 2, and 5 for synthetic data generation, data validation, and retrieval augmented generation (RAG) reliability assessment, respectively. The Army needs to be able to adopt large language models (LLMs) and LLM integrated systems (LLM-IS) in a manner that is secure and maintains user trust. Major barriers to user trust include the issues of LLM vulnerabilities and validation of the systems, so that end users can be confident in the LLM’s generated outputs. LLMs exhibit vulnerabilities unlike others observed in traditional information technology settings, owing to the unique nature of their data processing. This calls for innovative strategies for ensuring their cybersecurity readiness. Likewise, modern language processing techniques such as RAG must be verified before they can be properly deployed, but there is no current gold standard for such validations.Our primary focus in this project is on securing user trust in the adoption of LLMs and LLM-IS which is a key objective for both Project Linchpin and the Army’s Digital Transformation Strategy. Our key Army goals includes leveraging Zero Trust principles by establishing a standardized framework for cybersecurity evaluations of LLM and LLM-IS, in addition to validation strategies of RAG implementations.