The Pentagon is developing plans to allow AI companies to train military-specific versions of their models on classified data inside secure, accredited data centers. Currently, models like Anthropic's Claude are used in classified settings to answer questions but do not learn from the data they process. Training on classified data is expected to improve model accuracy for military tasks such as target identification and intelligence analysis. Key security risks include the potential for sensitive information—like the identity of operatives—to be inadvertently surfaced to unauthorized users within the military. Experts note that while leakage to the public internet is manageable, controlling information flow between different military departments with varying clearance levels is a harder problem. The initiative is part of a broader Pentagon push to become an 'AI-first warfighting force,' with existing agreements in place with OpenAI and xAI to operate models in classified environments.
Sort: