Cyber Command wants AI vendors to be replaceable

The Pentagon's AI fight is starting to look less like a contest to pick one blessed model provider and more like a scramble to avoid getting trapped by any of them. U.S. Cyber Command, the military command that runs offensive and defensive cyber operations, says it is building the plumbing to swap between AI systems as better ones appear, including open-source models from politically awkward places if they prove useful.

That makes this more than a small budget story. Axios reported that Brig. Gen. Reid Novotny, who leads acquisition and technology for Cyber Command, wants infrastructure that can support "an open-source made-in-China model" or "something very boutique" because operators need agility, not vendor loyalty. If that holds in procurement, value starts to move away from the model itself and toward the orchestration layer, the software and controls that decide which model gets used for which task.

The fresh trigger is concrete, if still modest. Axios reported that 2026 is the first year Cyber Command has dedicated funding for AI programs. DefenseScoop reported that the command's fiscal 2026 budget request includes $5 million for a new AI project and says the money would support 90-day pilot cycles inside the Cyber National Mission Force, the unit that carries out cyberspace operations against threats to the United States.

The amount is small beside the Pentagon's broader technology budget. But the design matters. According to DefenseScoop, Cyber Command wants AI across five categories: vulnerabilities and exploits, network security and monitoring, predictive analytics, persona and identity, and infrastructure and transport. That is not one model doing one demo. It is an attempt to wire AI into a stack of operational tasks where different tools may be better at different jobs.

That also helps explain why Novotny's quote matters. If Cyber Command expects one model to dominate every cyber task, there is no reason to emphasize switchability. If it expects capability to move fast, access to stay politically messy, and specialized models to keep appearing, then model choice becomes a workflow problem. The winner is not just the lab with the smartest model. It is whoever controls the layer that routes requests, manages permissions, logs outputs, and lets operators swap tools without rebuilding the whole system.

There is early evidence that the government's own behavior is already pushing in that direction. Axios reported that Anthropic restricted access to its offensive cyber model Mythos Preview to about 40 organizations. Axios also reported that the Cybersecurity and Infrastructure Security Agency did not have access even while the National Security Agency and the Commerce Department did. In other words, the constraint is not only raw model performance. It is who can use which system, under what terms, and with what political baggage.

That access problem gets sharper because the models themselves are becoming more capable, at least by the companies' own account. Anthropic said Mythos Preview had already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser, and said it was committing up to $100 million in usage credits plus $4 million in donations through Project Glasswing. Those claims come from Anthropic, not an independent public benchmark, and that caveat matters. Still, they help show why military buyers would want the freedom to switch quickly if one model pulls ahead on a specific task.

The skeptical case is straightforward. The roadmap is older than the current news cycle, the budget line is only $5 million, and the strongest new quote comes from one Axios interview. Cyber Command's own 2024 AI roadmap, according to a U.S. Cyber Command news release, outlined more than 100 activities, more than 60 pilot projects, and 26 new initiatives. Big roadmaps are cheap. Operational follow-through is not.

But even that caveat points at the real pressure. If Cyber Command is serious, it is building against a future where no one can safely bet on a single AI supplier, not even in national security. If it is not serious, then the Pentagon is still talking about agility while relying on ad hoc access to models controlled by private labs. The next thing to watch is whether Cyber Command's pilots produce visible procurement choices, security controls, or deployment rules that treat model vendors as replaceable parts instead of strategic destiny.