Pentagon plans to let AI companies train models on classified data

The US Department of War is working to set up secure environments where AI companies can train their models on classified data. Until now, models were only allowed to read classified data, not learn from it.

The Pentagon is discussing plans to let generative AI companies train military-specific model versions on classified data, MIT Technology Review reports, citing a US defense official.

Some AI models from major providers are already deployed in classified environments - for example, to analyze targets in Iran. But these models only answer questions without actually learning from the classified data. Training on it would be a significant leap: sensitive intelligence like surveillance reports or battlefield assessments would be embedded directly into the models themselves.

According to MIT Technology Review, this is the first known indication that LLM makers like OpenAI or xAI could train their models directly on classified data. Previous contracts only covered older computer vision models and commercially available data.

Accredited data centers would serve as secure training grounds

Training would take place in accredited data centers where a copy of an AI model is combined with classified data. The Department of War would retain ownership of the data, and AI company personnel would only gain access in rare cases with appropriate security clearance. Before using classified data, the Pentagon plans to first test how well models perform with unclassified data like commercial satellite imagery, the official said.

The Pentagon has been pushing AI adoption since a January memo from Defense Secretary Pete Hegseth that aims to turn the military into an "AI-first warfighting force."