Pentagon Plans Secure Environments for AI Companies to Train on Classified Data
The Pentagon is discussing plans to establish secure environments where generative AI companies can train military-specific model versions on classified data — including surveillance reports and battlefield assessments — going well beyond access to actively embedding classified intelligence into commercial model weights. A defense official confirmed the plans to MIT Technology Review on March 17, representing a significant policy escalation distinct from the existing Anthropic supply-chain risk designation and the OpenAI AWS deal. The arrangement would bring AI firms into unprecedented proximity with classified intelligence and creates novel risks: sensitive military knowledge permanently embedded in weights that could later be accessed via adversarial prompting or model extraction.
↳ Follow the thread