Defense & gov LLM moves

The Defense Intelligence Agency has put its own large language model, ChatDIA, onto the top-secret Joint Worldwide Intelligence Communication System as the Pentagon pushes more generative artificial intelligence into classified work. (breakingdefense.com) Breaking Defense reported on April 10 that ChatDIA was the first generative artificial intelligence chatbot to run on that network. Major General Robert Kinney said the system was first deployed on a smaller scale in fall 2025 and then expanded in December 2025. (breakingdefense.com) Federal News Network reported on April 13 that ChatDIA is part of a broader Defense Intelligence Agency push to scale artificial intelligence through a new Digital Modernization Accelerator, a permanent hub-and-spoke organization stood up in March 2026. The agency said the accelerator centralizes governance, funding and technical expertise while mission teams across directorates and combatant commands develop use cases. (federalnewsnetwork.com) A large language model is a system that predicts the next word in a sequence, like autocomplete trained on vast amounts of text. On a classified network, that lets analysts ask questions in plain English and search, summarize or draft against sensitive data without moving it onto public systems. (media.defense.gov) The Pentagon has been building toward this since August 10, 2023, when Deputy Secretary of Defense Kathleen Hicks created Task Force Lima under the Chief Digital and Artificial Intelligence Office to assess and guide generative artificial intelligence and large language model use across the department. The memo said the tools were growing in “popularity, capability, and impact” and required a coordinated approach. (media.defense.gov) The industry side moved this week too. OpenAI said on April 14 that it is expanding its Trusted Access for Cyber program to “thousands of verified individual defenders” and “hundreds of teams” and is starting with GPT-5.4-Cyber, a version of GPT-5.4 tuned for defensive cybersecurity work. (openai.com) OpenAI said the cyber model arrives ahead of “increasingly more capable models” planned over the next few months. The company’s main GPT-5.4 model was released on March 5 with a 1 million-token context window and positioning for coding, computer use and tool-based professional work. (openai.com 1) (openai.com 2) Microsoft also updated its in-house model lineup. The company introduced MAI-Image-2 in public preview last week and followed on April 14 with MAI-Image-2-Efficient, which Microsoft said is up to 22 percent faster and four times more efficient than MAI-Image-2 when normalized by latency and graphics processor use. (techcommunity.microsoft.com 1) (techcommunity.microsoft.com 2) Government agencies have been trying to keep these systems inside controlled environments rather than sending sensitive work to consumer chatbots. FedScoop reported in 2024 that Microsoft and Palantir were already adapting artificial intelligence and data tools for intelligence and defense agencies on classified networks, and the State Department said in 2025 that secure access to proprietary material such as diplomatic cables was central to its own generative artificial intelligence plans. (fedscoop.com 1) (fedscoop.com 2) The immediate question is not whether defense agencies will use large language models. It is how quickly they can move them onto secure networks, limit who can use them, and prove they help analysts without exposing classified data or lowering the standard for intelligence work. (federalnewsnetwork.com) (openai.com)