AI Distillation Erodes Pentagon's Model Advantage Over China

The Pentagon's AI Edge Faces a New Threat

The Department of Defense's shift toward AI-powered warfighting systems has created an unexpected vulnerability: adversaries no longer need to hack classified networks to steal military AI capabilities. Instead, they can simply clone the publicly available frontier models that underpin advanced defense systems through a technique called distillation.

From Project Maven's intelligence fusion to Anduril's Lattice sensor networks, the military's most sophisticated platforms rely on commercial AI models from companies like Anthropic, Google, and OpenAI. While the U.S. has implemented strict chip export controls to limit adversaries' compute capacity, these hardware restrictions cannot prevent the leakage of model intelligence through application programming interfaces.

Why It Matters

The gap between top U.S. and Chinese AI models has narrowed dramatically—from 17 percent in 2023 to just 2.7 percent on key benchmarks in 2026, according to the Stanford AI Index. This convergence threatens the technological asymmetry the Pentagon depends on for operational superiority, even as billions flow into AI-first defense systems.

How Distillation Bypasses Export Controls

Distillation allows a less capable "student" model to replicate a sophisticated "teacher" model's reasoning by learning from its outputs. Stanford researchers demonstrated this in 2023 by reproducing Meta's Llama model for $600—compared to the original $82,000 training cost.

Anthropic recently accused Chinese labs DeepSeek and Moonshot AI of generating millions of API calls from fraudulent accounts to systematically extract capabilities from their Claude models. These firms, founded only in 2023, rapidly deployed frontier-class systems at nearly 90 percent operational discount, according to details first reported by War on the Rocks.

While the U.S. controls roughly three-quarters of global high-end AI compute capacity versus China's 15 percent, distillation allows adversaries to bypass their hardware disadvantage. Every API query to a frontier model effectively leaks costly intelligence to anyone with a subscription.

A Two-Pronged Defense Strategy

Experts propose embedding Chief Digital and Artificial Intelligence Office liaisons directly inside frontier tech companies to gain early visibility into breakthrough capabilities months before public release. When significant advances emerge, the Pentagon would negotiate exclusive access windows—compensated through an "overmatch premium"—before commercial rollout triggers inevitable distillation.

This approach builds on the recent executive order establishing frameworks for early federal access to covered frontier models for up to 30 days before public release. The temporal advantage would compound across rapid four-to-six-month model release cycles, keeping the U.S. one generation ahead.

But early access alone provides only temporary reprieve. The Pentagon must also establish high-velocity refinement pipelines that transform secured models into specialized operational assets using theater-specific data, automated safety protocols, and systemic integration—at a pace adversaries cannot match through distillation alone.

The Clock Is Ticking

The U.S. military's "out-compute" strategy depends on maintaining model superiority. As Michael Kratsios, director of the White House Office of Science and Technology Policy, warned in a recent memo, APIs currently serve as an "unprotected pipeline" for American intellectual property. Without strategic intervention to manufacture time advantages and accelerate safe adoption, the Pentagon's AI-first transformation risks building on a foundation adversaries can replicate at will.

These findings and recommendations were detailed in an analysis published by War on the Rocks.