About 1.7 million of the Defense Department’s roughly 3 million personnel have now used GenAI.mil, the department’s enterprise generative AI platform, and around 500,000 of them use it heavily, a senior official said on 22 September at the DefenseTalks conference. The next step, he said, is to take frontier models beyond unclassified networks.
“We’re going to proliferate frontier models, not just on our controlled unclassified network,” said James Mazol, deputy under secretary of defense for research and engineering, DefenseScoop reported. He said the department would extend them to secret, top-secret and Special Access Program networks.
/ 01What happened
GenAI.mil launched in December for service members, civilians and contractors. The Army, Navy, Marine Corps, Air Force and Space Force have all adopted it as their preferred enterprise AI system. Mazol said that when the current administration began, about 80,000 department personnel were using generative AI, mostly homegrown models. Earlier this year the workforce used the platform to create 100,000 agents.
The model line-up has also widened. At the end of August the Pentagon added approved versions of OpenAI’s ChatGPT and xAI’s Grok alongside Google’s Gemini, DefenseScoop reported on 31 August. Press releases indicated all three had cleared Impact Level 5, the highest authorisation for non-public, sensitive unclassified data. A department official said it “will continue to build an architecture that prevents AI vendor lock.”
Not everyone on the panel sounded only optimistic. Cameron Stanley, the Pentagon’s chief digital and AI officer, said: “AI is getting put into everything. That scares me, mainly because of capacity.” He warned that compute will be hard to come by across industry and urged the department to ensure its models are “fit for purpose.”
/ 02Why it matters
Adoption at this scale is a real achievement, and it changes the question. With 1.7 million users and three frontier model families, the department is no longer testing whether people will use generative AI. It is deciding what work it should be trusted with, and on which networks.
Moving to classified networks raises the stakes in three ways. The data is more sensitive, so access controls and marking have to carry through every answer. The work is closer to operations, so a wrong answer has larger consequences. And the users are analysts and planners whose products are themselves cited and acted on, so an unsupported claim can propagate. Stanley’s phrase, “fit for purpose,” is the right test, and it is a different test for a policy memo draft than for an intelligence assessment.
The multi-model approach adds a further wrinkle. If a user can ask the same question of three models, the department needs a way to compare their answers on something firmer than fluency. Avoiding vendor lock-in is sensible procurement. It also means that whatever makes answers trustworthy cannot live inside any one vendor’s model.
/ 03Grounding that survives a model swap
The durable layer in a multi-model architecture is the data and the meaning attached to it. If the department’s knowledge, from equipment and units to facilities, regulations and reporting, is held in a governed knowledge graph, any model can be pointed at it, and any answer can be checked against it. Swapping Gemini for Grok, or adding the next model, does not change what counts as a fact or where it came from.
On classified networks this matters even more. A graph can carry classification and releasability as properties of each statement, so an answer assembled from several sources inherits the right markings by construction rather than by hope. It can also record provenance, so a reviewer can see which reports support a claim before it goes into a product. Agents built on that layer are easier to accredit, because their inputs are inspectable.
/ 04What to watch
- Which models reach which networks first, and under what accreditation.
- Compute. Stanley’s capacity warning will shape how quickly classified deployment can grow.
- Agents on GenAI.mil. With 100,000 already created, the department will need rules for which agents can touch which data.
- Evaluation. How the department defines “fit for purpose” for different mission areas, and whether it publishes the criteria.
With three model families behind one portal, the department should stop asking which model is most trustworthy and start asking which answers can be checked. That property has to live below the models, because the models will be swapped.
On a secret or top-secret network, a governed graph carries what a model cannot: each statement with its source report, its classification and its releasability. An answer assembled from five statements inherits the most restrictive marking by construction, and a reviewer can see which reports support each claim before it enters a finished product. Put the same question to Gemini, ChatGPT and Grok over that graph and the comparison stops being about fluency: the answers either rest on the same evidence path or they do not. Deterministic reasoning over the graph keeps the grounding identical whichever model writes the prose, and the whole stack runs behind the network boundary, where classified data has to stay.
How we’d model it
Statement—hasClassification→Classification markingStatement—sourcedFrom→Intelligence reportModel answer—assembledFrom→StatementAgent—authorizedFor→Network enclave
- GenAI.mil attracts about half a million 'power users' as Pentagon pushes forward with frontier models — DefenseScoop defensescoop.com ↗
- Grok and ChatGPT join Gemini in Pentagon's enterprise genAI portal — DefenseScoop defensescoop.com ↗
See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.