The Pentagon is planning for AI companies to train on classified data, defense official says

The Pentagon is planning to create secure facilities where AI companies like OpenAI and xAI can train military-specific AI models directly on classified data, a development that would embed sensitive intelligence into the models themselves. Defense Secretary Pete Hegseth issued a memo in January spurring the acceleration, as the department pursues an "AI-first" warfighting agenda amid escalating conflict with Iran. Training on classified data would make models more accurate for military tasks but introduces significant security risks—classified information embedded in models could potentially leak to unauthorized Pentagon departments or personnel. The Pentagon intends to first evaluate how effective models trained on non-classified data like commercial satellite imagery perform before proceeding with classified training. Palantir has already built secure infrastructure for asking AI questions about classified topics, but using these systems for training rather than querying represents a new security challenge the department has not yet fully implemented.