The Pentagon is pressing the leading AI companies to make their models available on classified military networks, and to do so with fewer of the usage restrictions the companies normally impose. Reuters reported on February 11, citing people familiar with the matter, that Pentagon Chief Technology Officer Emil Michael told technology executives at a White House event that the military wants AI models available on both unclassified and classified domains.
An official, speaking anonymously, told Reuters the department is “moving to deploy frontier AI capabilities across all classification levels.” Reuters could not determine how or when chatbots would be deployed on classified networks.
/ 01What happened
Most of the custom tools AI companies have built for the military run only on unclassified networks. The largest of these is GenAI.mil, the department’s enterprise platform. According to Reuters, OpenAI agreed to let the military use its tools, including ChatGPT, on that unclassified network, which is rolled out to more than three million Defense Department employees, and removed many of its usual user restrictions while keeping some guardrails. OpenAI said the arrangement covers unclassified use only and that expanding it would need a new or modified agreement. Google and xAI have reportedly struck similar deals.
Reuters reported that Anthropic is the only company whose model is currently available in classified settings, through third parties, and that the government remains bound by Anthropic’s usage policies there. Talks with Anthropic have been more contentious: its executives reportedly do not want the technology used to target weapons autonomously or for domestic surveillance in the United States. An Anthropic spokesperson told Reuters the company is “in productive discussions” with the department.
Pentagon officials, for their part, have argued they should be able to deploy commercial AI tools however they see fit as long as the use complies with U.S. law. The Decoder’s summary of the reporting notes that the companies involved include OpenAI, Anthropic, Google and xAI.
/ 02Why it matters
Classified networks are where the consequential work happens. Reuters notes they are used for sensitive tasks that can include mission planning and weapons targeting. Moving general-purpose models there is not just a procurement change; it puts probabilistic text generation closer to decisions with lethal and strategic consequences.
Reuters also reported that AI researchers warn that errors, including fabricated information, could have deadly consequences in those settings. That concern does not depend on where one stands on the usage-policy dispute. Whoever sets the rules of use, the model still has to be right, and when it is wrong, someone has to be able to find out why.
/ 03Accuracy and traceability on the high side
On an unclassified network, a bad chatbot answer usually costs time. On a classified network supporting planning or intelligence analysis, a fabricated detail that slips into a product can shape a decision. The usual mitigation — a human checks the output — only works if the human can see where each claim came from. A fluent paragraph with no citations forces the analyst to redo the research to trust it.
There is also an architectural point. Classified environments are disconnected from the commercial internet by design. Models deployed there have to run inside the enclave, with the data that lives inside the enclave. That favours systems that keep the organisation’s knowledge in a structured, inspectable form — entities, relationships, sources, and the provenance of each — so that a model of any size can reason over it and every answer can be traced to a specific document or record. It also makes it easier to change models later without rebuilding everything, which matters when vendor relationships are as unsettled as this week’s reporting suggests.
On a classified network, the question is not only who may use the model, but whether every answer can be traced to a source an analyst can check.
/ 04What to watch
- Contract terms. Whether agreements for classified use carry usage limits, and how they are written, will set a precedent for the whole sector.
- Accreditation. Running frontier models in classified enclaves requires accreditation and on-premises infrastructure. The timeline Reuters could not establish will depend on that work.
- Evidence requirements. Watch for any requirement that AI outputs used in planning or analysis carry source citations or audit trails. That would change which systems qualify.
The usage-policy fight gets the headlines, but on the high side the harder requirement is evidential. A planning staff cannot brief a course of action whose supporting facts came from a model’s memory.
Model the enclave’s holdings as a graph: units, capabilities, terrain and threat reporting, each node tied to the record that asserted it and its classification. A course of action then links to the capabilities it requires, the units that provide them and the intelligence assumptions it rests on. When a report is withdrawn or downgraded, every course of action that depended on it is found by traversal rather than by memory. Because the graph sits outside any one model, models can be swapped as vendor agreements change without losing the analytic record. Aktiver is built to deploy entirely inside a perimeter, and its regional-conflict build fused six sources into 720 records, each traceable to the document that produced it.
How we’d model it
Course of action—requires→CapabilityCapability—providedBy→UnitCourse of action—assumes→Intelligence reportIntelligence report—hasClassification→Classification level
- Exclusive: Pentagon pushing AI companies to expand on classified networks, sources say — Reuters (via Yahoo News) yahoo.com ↗
- Pentagon pushes AI companies to deploy unrestricted models on classified military networks — The Decoder the-decoder.com ↗
See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.