Pentagon Orders AI Pilot to Help Manage Military Secrets Within Six Months

An AI-assisted system could become the Pentagon’s authoritative digital reference for decisions about which military information must remain secret. The first step is a small-scale pilot: Deputy Defense Secretary Steve Feinberg issued a memorandum Monday directing deployment within six months, according to DefenseScoop, which reviewed the memo. Broader use depends on the trial succeeding.

The pilot will use the Automated Classification Management Environment, or ACME, a suite of AI-aided tools developed by the Air Force. It is designed to modernize security classification—the process of deciding how sensitive information is categorized—and the management of that information.

The Air Force’s top civilian official will oversee the pilot as its executive agent. A service spokesperson told DefenseScoop it would comply with the directive. The Pentagon did not immediately answer the outlet’s questions about which underlying AI models the trial would use.

A manual system under pressure

Feinberg’s memo argues that outdated classification and declassification procedures create incorrect classifications, obstruct cooperation between systems and organizations, and trap useful data in isolated channels. It also says delayed release of critical information erodes public trust.

An August request for information described the existing process as manual and labor-intensive. Hundreds of officials authorized to initially classify information use aging programs, the notice said, contributing to outdated documentation, high error rates, frequent over-classification and degraded data integrity. It also identified a roughly 140-million-page hardcopy backlog.

Josh Wallin, a fellow studying AI in the Center for a New American Security’s defense program, cautioned against complete trust in an evolving system.

Faster sharing, faster mistakes

Wallin told DefenseScoop that automation could help correct human mistakes and reduce information bottlenecks. Over-classification can delay reviews and keep information from organizations that need it, including coalition partners. Under-classification carries the opposite danger: exposing a military capability or plan to an adversary.

He identified several ways faster processing could introduce risk:

• Classification errors could occur at greater speed and scale, allowing sensitive information to be shared before someone catches the mistake.
• Weak cybersecurity among allies or partners could give adversaries another route to information, even if U.S. systems remain protected.
• Separately declassified pieces of information could collectively reveal something sensitive, an aggregate effect that faster declassification might accelerate.

Wallin also questioned the pilot’s apparent brevity and how human oversight would work if ACME became the definitive reference for original classification decisions. He called for procedural checks to manage inevitable mistakes in both human and AI-driven processes, and said human oversight would need to persist permanently as the system and circumstances change.