Pentagon Plans AI Training on Classified Data
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Pentagon is developing secure data centers to allow AI firms to train models on classified information, a move that would embed sensitive intelligence into AI systems for military use
- Anthropic's AI models are already used in classified settings for tasks like analyzing Iranian targets, though not yet trained on classified data
- OpenAI and xAI have reached agreements with the Pentagon to operate their models in secure environments as part of a broader 'AI-first' warfighting strategy
- AI models trained on classified data could become more accurate in military applications, according to a U.S. defense official, but would require strict access controls and security accreditation
- AI company personnel may access classified data only if they hold appropriate security clearances, though the Pentagon retains ownership of the data used in training
- Pentagon plans to first assess model performance using unclassified data, such as commercial satellite imagery, before authorizing training on sensitive material
Why it matters: This shift brings private AI firms into direct contact with classified military intelligence, increasing model effectiveness but also expanding the attack surface for data breaches. The defense official’s emphasis on evaluating nonclassified training first suggests a cautious approach to a high-stakes integration that could redefine battlefield decision-making.
