Eighteen Weeks and Still Can't Read a Secret (liminaldr.substack.com)

The Pentagon's GenAI.mil initiative, launched in December 2025, has faced significant challenges in accessing classified networks after eighteen weeks of operation. Despite being designed for Defense Department personnel, the AI system is unable to interact with classified data due to its reliance on commercial cloud infrastructure, which is air-gapped from classified networks. This situation highlights a broader concern within the AI/ML community regarding the secure integration of AI tools with sensitive government information. Existing models in the intelligence community, such as GAIA, have demonstrated that classified AI capabilities can be developed internally, yet the essential missing component is a secure ingestion layer that allows for safe and compartmentalized access to classified documents. As AI integration with classified data unfolds, the emphasis is on the development of Retrieval Augmented Generation (RAG), which enables models to retrieve relevant context from secure databases without compromising sensitive material. Companies like Ask Sage and Unstructured are pioneering this technology to create secure interfaces that separate AI models from the classified data they process. The critical issue remains whether these private vendors can maintain stringent security standards, especially as their architectures become essential for accessing the government's most sensitive information. This ongoing dependency on external vendors raises questions about security oversight and the potential risks of misaligned incentives should these companies be acquired or change their operational priorities.