The Pentagon is considering the establishment of secure facilities where generative AI companies can develop military-oriented models utilizing classified data. Currently, AI systems like Claude from Anthropic are employed in classified scenarios, such as analyzing targets. However, the prospect of training these models on sensitive data introduces significant security challenges. This would mean that critical intelligence, such as surveillance data, could be integrated into the AI systems, increasing the exposure of classified information to these companies.
A defense official indicated that training AI on classified data could enhance the accuracy and effectiveness of these models for specific military tasks. This initiative comes amid rising demand for advanced AI capabilities, with the Pentagon collaborating with firms like OpenAI and xAI to utilize their technologies in secure environments. Training would occur in accredited data centers, with AI personnel potentially accessing classified information under strict security protocols.
Before proceeding, the Pentagon plans to assess the performance of models trained on unclassified data. Historically, the military has utilized AI for image analysis and has contracted firms to develop models for government applications. However, the idea of training AI directly on classified data marks a significant shift in the military's approach to AI technology, raising concerns about information leaks and security risks.