The Pentagon is exploring initiatives to establish secure environments for generative AI firms to develop military-specific iterations of their models using classified data, MIT Technology Review has discovered.
AI models such as Anthropic’s Claude are already deployed to respond to inquiries in classified contexts; uses include scrutinizing targets in Iran. However, permitting models to train on and absorb classified data would mark a significant advancement that could introduce distinct security vulnerabilities. It would imply that sensitive intelligence, like surveillance documentation or battlefield evaluations, could be integrated into the models, bringing AI companies into closer proximity to classified information than previously experienced.
Training AI models with classified data is anticipated to enhance their precision and efficiency in certain functions, according to a U.S. defense official who spoke on condition of anonymity to MIT Technology Review. This development comes as the demand for more capable models surges: The Pentagon has made arrangements with OpenAI and Elon Musk’s xAI for the operation of their models in classified environments and is executing a new strategy to transform into an “AI-first” combat force as tensions with Iran intensify. (As of publication time, the Pentagon had not commented on its AI training initiatives.)
Training would occur in a secure data center accredited to accommodate classified governmental projects, where a duplicated version of an AI model is linked with classified data, as per two individuals familiar with the operational procedures. While the Department of Defense would maintain ownership of the data, personnel from AI companies might, in rare instances, access the data if they possess the necessary security clearance, the official stated.
Prior to sanctioning this new training, the official indicated, the Pentagon plans to assess the accuracy and efficiency of models trained on unclassified data, such as commercially available satellite imagery.
The military has historically employed computer vision models, an earlier type of AI, to detect objects in images and footage collected from drones and aircraft, while federal agencies have awarded contracts for companies to train AI models on such materials. Furthermore, AI firms developing large language models (LLMs) and chatbots have produced iterations of their models optimized for governmental functions, including Anthropic’s Claude Gov, designed to work across multiple languages and in secure settings. The official’s remarks, however, represent the first sign that AI companies developing LLMs, such as OpenAI and xAI, might train government-specific versions of their models directly using classified data.
Aalok Mehta, director of the Wadhwani AI Center at the Center for Strategic and International Studies and former head of AI policy at Google and OpenAI, asserts that training on classified data, rather than merely answering queries about it, could pose new risks.
The foremost risk, according to him, is that classified information which these models are trained on might be revealed to anyone utilizing the model. This would pose challenges if numerous different military branches, each with varying classification levels and information access requirements, were to utilize the same AI.
“You can envision, for instance, a model with access to sensitive human intelligence—like an operative’s identity—leaking that information to a segment of the Defense Department not authorized to access it,” Mehta explains. This could create a security threat for the operative, one that is hard to completely mitigate if a specific model is employed by multiple groups within the military.
Nonetheless, Mehta mentions that it’s not overly challenging to confine information from being accessed by the wider public: “If you configure this correctly, there will be minimal risk of that data appearing on the general internet or reverting back to OpenAI.” The government possesses some of the necessary infrastructure already; the security firm Palantir has secured significant contracts for creating a secure environment through which officials can inquire AI models about classified issues without relaying the information to AI companies. However, utilizing these systems for training represents a novel challenge.
The Pentagon, prompted by a memo from Defense Secretary Pete Hegseth in January, has been hastily working to integrate more AI. It has been employed in combat situations, where generative AI has compiled lists of targets and suggested which ones to prioritize, as well as in more bureaucratic capacities, such as drafting contracts and reports.
There are numerous functions currently conducted by human analysts that the military may wish to train prominent AI models to execute, which would necessitate access to classified information, Mehta notes. This could include learning to identify subtle indications in an image in the same manner an analyst would or linking new data with historical context. The classified data could be sourced from the immense volumes of text, audio, images, and videos, in various languages, that intelligence agencies gather.
It’s exceedingly challenging to identify which precise military tasks would require AI models to train on such data, Mehta warns, “because undoubtedly the Defense Department has substantial incentives to keep that data confidential, and they don’t want rival countries to be aware of the exact capabilities we possess in that domain.”
If you possess information regarding the military’s application of AI, you can confidentially share it via Signal (username jamesodonnell.22).
