How Federal Agencies Are Building Secure, Mission-Ready GPTs

Federal agencies are developing and deploying their own custom generative pre-trained transformers (GPTs) as they adopt artificial intelligence to automate administrative tasks, write code and augment human analysis. Systems like the Air Force Research Laboratory’s Non-Classified Internet Protocol Generative Pre-training Transformer (NIPR GPT) can help public servants do their jobs better, according to experts.

“[Generative AI] gives a non technical expert the ability to build simple applications very, very quickly. And for an engineer who knows what they’re doing, for stitching capabilities together, it’s an even larger game changer,” said Collen Roller, former AFRL senior computer scientist at the Air Force Research Laboratory and NIPR GPT lead, during the Defense IT Summit in February. “What I always like to say is, if you’re not using gen AI in 2025, you’re slow at your job.”

The acceleration in adoption follows the introduction of government-specific commercial offerings like OpenAI’s ChatGPT Gov, which allows agencies to self-host the technology within secure Microsoft Azure government clouds. Self-developed models, officials say, are essential for reducing administrative burdens and accelerating mission-critical tasks while handling sensitive information.

“Working on a classified network is like working in a barren wasteland. You are not able to access the open Internet. You certainly do not have large language models,” former Central Command CTO Schuyler Moore said during the NVIDIA Summit in Washington, D.C last year. “Your experience, or especially if you are a programmer, as you can imagine, is quite painful.”

CISA Builds Chatbot to Streamline Cyber Operations

National security leaders are focusing AI efforts on air-gapped, internal models that accelerate analysis of large, sensitive but unclassified datasets.

The Cybersecurity and Infrastructure Security Agency (CISA) has developed internal tools like CISAChat, which allows authorized personnel to interact with, summarize and search agency-created materials and internal threat intelligence reports. The chatbot works with Microsoft’s M365 Copilot to streamline administrative work and has passed agency privacy and security reviews, CISA CIO Bob Costello told GovCIO Media & Research in May.

“CISA staff can now generate content, summarize documents, analyze data and manage schedules more efficiently, all while adhering to agency security protocols,” Costello added.

Pentagon Turns to GPTs to Boost Mission Readiness

At the Pentagon, generative AI is transforming employee experience. AFRL launched NIPR GPT in 2024 to serve Air Force personnel. The secure, chatbot-style application was developed by a small team to further the Air Force’s modernization effort, according to Roller.

“For my team, we’re small. We’re literally three guys, and we built around 12 web apps,” Roller said during a GovCIO Media & Research DevSecOps Working Group meeting last year. “We partner with the larger software factories. We got a really tight relationship with Platform One and Cloud One.”

NIPR GPT initially provided Airmen, Guardians, civilian employees and contractors a sanctioned environment to experiment with generative AI. Roller said it later expanded to other War Department entities due to unexpected demand.

“We originally hypothesized that only people at AFRL were going to use this thing … We thought that a lot of people would just try to go to ChatGPT and circumvent the government systems,” Roller said during the NVIDIA event. “We were completely shocked that that was not the case … The demand for gen AI in the [War Department] is real. It’s something that we collectively need to care about and focus on.”

The Army mirrored this approach with CamoGPT, an experimental generative AI platform from its AI Scholars Program. The tool enables users to upload documents, interact via natural language, and generate summaries, metrics and visualizations, according to DOW. The platform also supports custom tool integration and API access, making it adaptable for developer workflows, according to the Army.

Col. Isaac Faber, Army data scientist and former director of the Army’s Artificial Intelligence Integration Center, said soldiers trained at Carnegie Mellon University lead program development.

“We built a workforce development program … Go to Carnegie Mellon, get an AI master’s degree for two years, and then for the next two years, you come to me in the AI center and you work in our incubator,” Faber said at the NVIDIA event. “We’re chartered to make and experiment early-stage products.”

CamoGPT now supports tens of thousands of users across the Army, letting soldiers learn AI on their own terms, Faber said.

“It’s not an enterprise or fielded solution,” Faber said. “It’s an S&T solution that allows us to ask really hard questions like, ‘What are soldiers actually using it for when we put it in their hands?’”

At Central Command, CENTGPT  — built off NIPR GPT’s code — was developed and deployed on classified networks to support crisis operations. Following the outbreak of the Israel-Hamas war, CENTCOM faced a surge in meetings and reporting demands, according to Moore.

“This crushing weight descended on our staff,” Moore said during the NVIDIA event in 2024. “Suddenly, the requirement to participate, summarize and push all of this information … became overwhelming.”

While generic summarization proved less effective than hoped, CENTCOM found success in three key areas: code generation, machine-assisted disclosure and staff work augmentation.

“Our programmers were overjoyed,” Moore said. “In a minute of having a large language model available on a [classified] environment, we opened the complete aperture of what our programmers could do.”

NIH Harnesses AI to Advance Research and Efficiency

The National Institutes of Health (NIH) is leveraging AI to enhance research and enterprise-level access management. Leaders said at a September event that the agency is rolling out enterprise-grade tools for administrative and scientific tasks.

“We’re really trying to enable all of our staff to use those tools and explore how [AI] can impact their day-to-day activities,” said Nick Asendorf, science information officer at the National Heart, Lung and Blood Institute (NHLBI), adding collaboration is crucial to scaling AI adoption.

“We are sharing new applications that we develop, and we’re trying to learn when is something specific to an institute, or when is it a similar problem that three people around the table are experiencing at the same time,” said Asendorf.

NHLBI Chat launched in 2024 to allow NIH staff to safely explore generative AI tools, according to Asendorf.

“Really what we were trying to do is give our staff the ability to use this technology and see how it could impact them,” said Asendorf at the event. “Since our staff has [NHLBI Chat], we have been partnering across NIH to give this type of tooling to make them more efficient in their day-to-day job.”

NIH Center for Information Technology (CIT) Director Sean Mooney NIH highlighted the Science and Technology Research Infrastructure for Discovery, Experimentation, and Sustainability (STRIDES) Initiative, which harnesses existing infrastructure to help the NIH revolutionize tech modernization efforts. Within the program, the NIH has deployed Microsoft CoPilot and ChatGPT Enterprise for intramural research.

“Interactive chatbots that are incredibly helpful and and the benefits of parsing text and language, which up until these models were produced, have been very difficult to do,” said Mooney.

Generative AI Yields Tangible Impact Across Federal Missions

Generative AI is delivering measurable mission impact by streamlining administrative tasks, Roller said.

“My favorite use case is putting all my team’s meetings into NIPR GPT and creating a roll-up: what did I do today?” he explained.

Machine-assisted disclosure proved transformative at Central Command, letting users focus on critical analysis. “It can flag documents with high confidence for disclosure, freeing humans to focus on gray areas,” Moore said.

GPTs are also critical for workforce retention. “Digital-native staff are accustomed to commercial products,” Moore said. “If we cannot offer a comparable experience at work, people will leave.”