Aktiver — semantic AI platform Deterministic · traceable · behind your firewall
Insights / Defense

The Pentagon signs eight AI firms for its classified networks

On May 1, 2026, the Department of Defense announced agreements with seven technology companies to deploy their AI on its classified networks: Amazon Web Services, Google, Microsoft, NVIDIA, OpenAI, SpaceX and Reflection, a startup backed by NVIDIA. Oracle was added hours later, bringing the total to eight. The agreements cover the department’s Impact Level 6 and Impact Level 7 environments, the networks that handle secret and the most highly classified data.

Anthropic, whose Claude models had already been in use on classified networks through Palantir’s Maven toolkit, was not on the list. The company is in litigation with the department after being labelled a supply chain risk following a dispute over limits on military use of its models.

/ 01What the Pentagon announced

According to DefenseScoop, the department said the work will “streamline data synthesis, elevate situational understanding, and augment warfighter decision-making,” and tied it to its goal of an “AI-first fighting force.” Breaking Defense reported that some of the companies are already under contract while others are still finalizing details, and that the announcements did not say when the models would be available on classified networks or how much the companies will be paid.

The move extends GenAI.mil, the secure but unclassified generative AI platform the Pentagon launched in December. Per the department, more than 1.3 million personnel have used it, generating tens of millions of prompts and hundreds of thousands of agents in five months. The Associated Press reported the Pentagon’s claim that some tasks have gone “from months to days.”

Pentagon CTO Emil Michael framed the mix of vendors as deliberate. Speaking to CNBC, as quoted by Breaking Defense, he said “it’s irresponsible to be reliant on any one partner,” and described a “diversity of supply” that spans open-source models, proprietary model companies and infrastructure providers such as Microsoft and AWS. The AP noted that NVIDIA and Reflection, both makers of open models, are new to classified work, while Amazon and Microsoft have long operated in classified settings.

/ 02Why it matters

Two things change at once. First, frontier models move from unclassified desktops to the networks where operational planning and intelligence analysis actually happen. Second, the department is committing to a multi-model estate rather than a single provider. Lauren Kahn of Georgetown’s Center for Security and Emerging Technology told DefenseScoop that access to multiple models helps guard against vendor lock-in and helps users learn how different systems behave, and that training on automation bias is a necessary next step.

The AP reported that at least one contractor’s agreement requires human oversight in certain situations and use consistent with constitutional rights and civil liberties, and that OpenAI says it secured similar assurances. The terms of most agreements were not published.

A multi-vendor posture is sensible procurement. It also means that the same question, put to two different models on the same classified network, may get two different answers. On an unclassified chat portal that is an inconvenience. In a targeting cell or an intelligence fusion centre it is a governance problem: which answer was used, what evidence supported it, and could anyone reconstruct the chain later?

/ 03The traceability angle

Language models are interchangeable at the interface and very different underneath. What should not be interchangeable is the knowledge they reason over. If the facts a commander relies on — units, locations, equipment, reporting sources, their reliability and their timestamps — live in a governed knowledge graph, then any model on the network can be pointed at the same structured record, and its outputs can be checked against it.

That design has three practical benefits in a classified, multi-model setting:

  • Model swaps stop being risky. If a vendor drops out or a better model arrives, the institutional knowledge stays in the graph, not in prompts and fine-tunes tied to one provider.
  • Disagreements become inspectable. When two models differ, each claim can be traced to the graph nodes and source documents behind it, instead of being settled by whichever output sounds more confident.
  • Automation bias has a counterweight. An analyst who can see the evidence chain behind a recommendation has something to check, which is the kind of habit the CSET comment points toward.

None of this is exotic. The Common Core Ontologies, which extend the Basic Formal Ontology, grew out of defence and intelligence data-integration work. What is new is the pressure: once frontier models are on IL6 and IL7 networks, the question of what grounds them stops being academic.

Swapping models should be easy. Swapping the facts they reason over should be impossible without a record.

/ 04What to watch

The announcements left the hard details open. Watch for when each model actually reaches classified networks, what the published or leaked contract terms say about human oversight and permitted uses, and how the Anthropic litigation resolves. Watch, too, for whether the department sets common evaluation and logging requirements across all eight vendors, which would matter more for accountability than any single model choice.

For programme offices and integrators, the practical step is to separate the knowledge layer from the model layer now. Whichever models end up on the network, the record of what they were told and why they answered as they did should belong to the mission, not the vendor.

How it unfolded
DEC 2025GenAI.mil platform launched[2]MAY 1, 2026Seven firms cleared forIL6/IL7[2]MAY 1, 2026Oracle added as eighth firm[1]MAY 3, 2026This postDEC 2025GenAI.mil platform launched [2]MAY 1, 2026Seven firms cleared for IL6/IL7 [2]MAY 1, 2026Oracle added as eighth firm [1]MAY 3, 2026This post
Aktiver’s take — Defense & intelligence

Eight vendors on IL6 and IL7 is a procurement answer to a vendor-risk problem. It leaves the harder question open: when two models give a fusion cell two different readings, which one went into the decision, and on what evidence?

In an ISR fusion workflow the graph sits under every model on the network. Collection reports, units, equipment and locations become typed records, each tied to its originating report, the source’s reliability rating and a timestamp. A model’s assessment is stored as a claim that cites those records, so an analyst can ask which reports place a battalion at a crossing and whether they are independent of each other. Because reasoning over the graph is deterministic, the same question returns the same answer and the same evidence whichever vendor’s model drafted the prose. Swapping a model then changes the wording of an assessment, not the record it rests on, and the whole stack stays inside the classified enclave.

How we’d model it

AssessmentIntelligence reportSource ratingUnitLocationModel outputModel versioncitesEvidencehasReliabilityobservedAtgeneratedByAssessmentIntelligencereportSource ratingUnitLocationModel outputModel versioncitesEvidencehasReliabilityobservedAtgeneratedBy
  • Assessment —citesEvidence→ Intelligence report
  • Intelligence report —hasReliability→ Source rating
  • Unit —observedAt→ Location
  • Model output —generatedBy→ Model version

See where this applies →

Go deeper

See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.

← All insights Subscribe via RSS → Book a demo →
Keep reading