The Pentagon is developing plans to create secure environments where AI companies like OpenAI and Anthropic can train military-specific versions of their models on classified data. This represents a significant shift from current practices where AI models answer questions in classified settings without learning from the data, potentially improving performance for defense tasks.

Training would occur in accredited secure data centers where copies of AI models pair with classified information. While the Department of Defense would retain data ownership, cleared AI company personnel might occasionally access sensitive intelligence. The Pentagon plans to first evaluate model accuracy using non-classified data before implementing this new approach.

Security expert Aalok Mehta warns that classified information could resurface to unauthorized users, creating risks for operatives if models are shared across military departments with different classification levels. The Pentagon, spurred by Defense Secretary Pete Hegseth's memo, has been rapidly integrating AI into combat and administrative roles as tensions with Iran escalate.