Pentagon AI Classification Pilot Moves Secrecy Decisions to

Deciding what the government must keep secret has long been work for specially designated human officials. A Pentagon AI classification pilot now aims to test whether software can become the department's reference point for those decisions. Deputy Defense Secretary Steve Feinberg signed a memo ordering a small-scale deployment within six months, built on the Air Force's Automated Classification Management Environment, known as ACME.

The memo sets a 180-day window for the executive agent to roll out a limited automated security classification capability. If the pilot meets its goals, the system would become the department's single digital reference for original classification decisions. Original classification is the step where information is first judged to require protection, so the pilot targets the front end of the secrecy process rather than the review of documents after the fact.

How the pilot is built

ACME is an existing Air Force system, which means the pilot starts from a tool the service already runs rather than a new platform. The memo's language points to a phased approach: prove the capability in a limited setting first, then decide whether it takes on a wider role. The sources do not name the vendors, budget or classification levels involved in the pilot, so the scope beyond the memo's six-month target remains unclear.

The stakes of the design are concentrated in one function. A single digital reference can speed up decisions and make them more consistent across components, but it also concentrates authority over what is classified in one system. An error in that system would propagate to every document that cites it, which is why the memo's conditional framing matters: the reference role depends on the pilot meeting its goals.

Reporting on the memo also describes a separate timeline for the same effort. Coverage from several outlets repeats the six-month deployment target, while the memo's own wording frames the rollout as a limited capability within 180 days. Both formulations point to the same near-term milestone, which gives program staff a fixed date to plan against. Some outlets refer to the department as the Department of War and to Feinberg as Deputy Secretary of War, the title used in the same announcement coverage, so the memo's authority runs through the top civilian tier of the building.

Where the pilot fits in the Pentagon's AI push

The classification pilot follows a steady expansion of AI inside the department. GenAI.mil has reached 1.7 million Pentagon users, and the platform is moving from unclassified tasks toward sensitive and eventually classified environments. Adoption at that scale raises the cost of failure, a concern that tracks directly with the move to automate classification.

A separate department strategy memo directs broad use of AI across the military, with the aim of placing leading models in the hands of roughly three million civilian and military personnel at all classification levels. Read together with the classification memo, the direction is clear: AI is being pushed into both the handling of secrets and the decisions about which secrets exist.

Funding has followed the same direction. Last summer the Pentagon gave xAI, OpenAI and Google $200 million apiece to develop self-directed AI systems capable of handling classified material with little need for ongoing human oversight. Those awards predate the classification memo, and the pilot is the first public step that puts AI on a formal classification role rather than on analysis or workflow support.

Targeting software shows the same pattern. The Maven Smart System combines sensor, intelligence, surveillance and command data for targeting support, and Claude models were integrated into Maven through Palantir's AI platform in late 2024. Anthropic later announced Claude Gov, a version that runs in classified environments. These deployments show that AI models already touch classified material, and the classification pilot would move that contact into the rules that decide what counts as classified.

Operational AI is also expanding in the space and missile domain. The Pentagon is seeking an AI system that can turn space, missile and intelligence data into threat assessments within seconds, and the department created an Autonomous Warfare Command, reported as Autowarcom, to expand AI and drone capabilities. Those programs share a demand for fast machine-generated judgments, which the classification pilot applies to information control.

Human judgment and oversight

The Congressional Research Service has examined the policy question that sits underneath automated decisions. Its analysis of the Pentagon-Anthropic dispute over autonomous weapon systems argues that meaningful human control does not require a person to operate every step by hand. It requires broader human involvement in deciding how, when, where and why a system is employed. Applied to classification, that standard would mean a human official still assesses the context and owns the decision, even when a model drafts the call.

Recent reporting has also shown what happens when AI output is trusted without enough checking. CNN reported in September 2026 that US military personnel had a close call after using AI to produce an intelligence report containing false information. The episode is the kind of failure a classification reference would need to catch, because a wrong classification decision would be harder to reverse than a wrong briefing.

What changes next

For program offices and contractors, the practical question is who validates the system's classification calls and how those calls are logged. The memo ties the reference role to meeting pilot goals, so the 180-day mark will show whether the department treats the tool as advisory or authoritative. Vendors with classified-work contracts, including those that already supply models to Maven and GenAI.mil, are positioned to compete for the follow-on work that a permanent reference would require.

Current reporting relies on the memo's description and on secondary coverage that repeats it. The memo itself is the primary source, and the department has not yet published pilot results, performance metrics or an oversight framework for the automated reference. Those documents will determine whether the pilot changes classification practice or remains a limited test.

Why this matters

The pilot tests whether software can hold a formal role in deciding what the government keeps secret, a responsibility that has rested with human officials. For contractors and defense technology buyers, the six-month timeline sets a near-term benchmark for AI in classified workflows. The result will show how far the Pentagon is willing to move AI from analysis into the rules that govern information control.

Sources

Pentagon-Anthropic Dispute over Autonomous Weapon Systems: Potential Issues for Congress | Congress.gov | Library of Congress

Photo by Mirella Callage on Unsplash

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.