type: timeline_event
On March 17, 2026, MIT Technology Review revealed that the Pentagon was planning a significant expansion of AI companies' access to classified intelligence data — moving beyond the current practice of running queries against classified databases to allowing companies to train military-specific AI models directly on classified information. The shift would embed surveillance reports, battlefield assessments, signals intelligence, and other sensitive material into the actual weights of foundation models, representing a fundamental change in how classified data interfaces with commercial AI systems.
The plan called for training to take place in accredited secure data centers meeting stringent government security requirements. OpenAI and Elon Musk's xAI had already signed agreements to participate in the program, positioning them to build specialized military AI models that would internalize classified knowledge rather than merely accessing it at query time. The distinction was technically significant: a model trained on classified data would carry that information in its parameters permanently, raising novel questions about data security, model portability, and the risk of classified information leaking through model outputs.
The reporting raised alarm among AI safety researchers and civil liberties organizations. Training on classified data would create models whose behavior was shaped by surveillance and intelligence information that no external auditor could review, making independent safety evaluation effectively impossible. The plan also deepened concerns about the merger of commercial AI companies with the intelligence apparatus — once model weights contained classified data, the companies building them would be permanently entangled with the national security state in ways that could not be easily unwound. Critics noted that the Pentagon was simultaneously blacklisting the one major AI company that had pushed back on unrestricted military use while fast-tracking classified data access for companies that had raised no objections, creating powerful incentives for the industry to abandon safety positions.