The Pentagon's AI Rush Is Creating a Security Weakness

The Department of War just announced on August 31, 2026, that it is expanding GenAI.mil with ChatGPT Mil and Grok for Government, adding more Corporate AI directly into the daily infrastructure of the American military. Both systems are accredited to handle Controlled Unclassified Information at Impact Level 5, and GenAI.mil has already reached more than 1.7 million unique users since launching in December 2025. This is not a small test anymore. The Department has openly called for an “AI-first” force and says these tools are meant to accelerate planning, policy, logistics, administration, acquisition, supply chains, and other routine work. The goal is to make AI normal, make it fast, and put it in the hands of as much of the force as possible. That may produce real advantage, but it also creates a basic security problem that every risk analyst at the NSA should be modeling every day. Rapid adoption means more users, more credentials, more stored projects, more data movement, more software connections, and more opportunities for someone on the other side to find a way in.

That matters because people ask questions, upload files, copy text, summarize material, save projects, and return to the same system day after day until the interaction feels ordinary. The more ordinary that interaction becomes, the easier it is for security discipline to erode around it. Mass adoption of conversational AI does not automatically violate OPSEC, but it places greater pressure on it. The Pentagon’s own doctrine warns that data aggregation, routine information systems, and ordinary organizational activity can create exploitable vulnerabilities. GenAI.mil is expanding all three at once.

America has already watched military enthusiasts repeatedly place restricted or sensitive material onto War Thunder forums simply to win arguments about weapons. The point is not that GenAI.mil is War Thunder. The point is that human beings are human beings. Give millions of people a tool built around frictionless conversation and then make that tool part of their normal workday, and some portion of them will eventually become too comfortable. Security rules, SOPs, training, and accreditation can help reduce the risk. None of them can predict all user behavior across millions of interactions. The earlier warning in Corporate AI as the Military’s Weakest Link was that normalization itself becomes the vulnerability due to the erosion of standards. The Department of War is now explicitly trying to make frontier AI the standard for daily operations.

The attack surface is getting larger whether each individual model is secure or not. More AI systems mean more software, more integrations, more credentials, more stored knowledge, and more places a hostile actor can probe. NIST’s own Generative AI Risk Management Profile explicitly warns that generative AI expands the available attack surface. DoD officials have separately told GAO that generative models could aggregate unclassified information in their training data and unintentionally output classified information. A hostile intelligence service does not need to defeat the entire Department of War.

The critical mistake intelligence and the military are making is believing that training can solve this problem. The Strange Consolation of AI Ethics dealt with the problem of simulated accountability, where institutional structures can provide reassurance without materially constraining the underlying risk. The same danger applies here. An IL5 accreditation matters, but it does not make 1.7 million users disciplined. A risk framework matters, but it does not guarantee that routine use will remain careful five years from now. Just because we haven’t seen a catastrophic breach from this military AI rollout yet doesn’t mean it won’t happen.

The United States should use AI where it creates a real military advantage, but rapid adoption should not be confused with stronger warfighting. The Pentagon should carefully measure AI adoption against its own procedures, identify the security burdens it creates, and mitigate those risks downstream. Every new AI capability creates another thing that has to be defended, and every layer of convenience puts more pressure on the oldest security system the military has, the judgment of the person behind the keyboard. America’s adversaries do not need to invent a superintelligence to exploit this. They only have to wait for discipline to erode.

Michael Aaron Cody is an independent theorist published in Physics Essays, Journal of Modern Physics, European Physical Journal Plus (Springer Nature), FPIF, Small Wars Journal, and RealClearScience. His work spans physics and defense policy. Find him on X at https://x.com/MichaelCody.

Originally published by Foreign Policy in Focus.