Deidentifying Personally Identifiable Information (PII) for LLM Ingestion and RAG Applications

With the AI race intensifying, state-sponsored cyber actors are increasingly targeting AI vendors to exploit vulnerabilities for strategic advantage. A recent breach in OpenAI’s Large Language Model servers and vector databases exposed sensitive personnel data and research, highlighting the growing risks of AI in defense.Retrieval Augmented Generation (RAG) will enhance decision-making processes, streamline operations, and improve mission planning, but Personally Identifiable Information (PII) is dispersed throughout military data, creating significant privacy and compliance risks. Manual redaction of this data is inefficient and prone to human error.Tonic AI, Inc.’s Tonic Textual is the solution to deidentify sensitive data before training on LLMs, paving the way for model optimization and scalable techniques for adoption of RAG.