Genetic AI models used in confidential environments can answer questions, but for now they do not learn from the data they see. This could change soon. The Pentagon discusses plans to create secure environments where genetic AI companies can train militarily specialized versions of their models with confidential data, as reported by MIT Technology Review.
See also: Anthropic vs. Pentagon: Support from Google and OpenAI
AI models such as Anthropic's Claude are already being used to answer questions in classified settings, including target analysis in Iran.
However, training models with confidential data would be a new development that presents unique security risks. This would mean that sensitive information, such as surveillance reports or battlefield assessments, would be embedded into the models themselves, bringing AI companies into closer contact with confidential data than before. Training versions of AI models with confidential data is expected to improve their accuracy and effectiveness on certain tasks, according to a U.S. defense official who spoke anonymously to MIT Technology Review.
This news comes as demand for more powerful models is high: the Pentagon has signed agreements with OpenAI and Elon Musk's xAI for operating their models in classified settings and is implementing a new agenda to become an “AI-first” warfighting force as the conflict with Iran escalates. Training will be carried out in a secure data center certified to host classified government projects, where a copy of an AI model will be combined with classified data.
Even though the Ministry of Defense will remain the owner of the data, staff from AI companies with appropriate security clearances can, in rare cases, access the data. Before this new training is allowed, the official said the Pentagon intends first to assess how accurate and effective the models are when trained on non‑classified data, such as commercially available satellite images.
See also: CMMC: Pentagon requires verified cybersecurity from contractors
The military has been using computer vision models, an older form of AI, for a long time to recognize objects in images and footage collected from drones and aircraft. Federal agencies have awarded contracts to companies to train AI models on such content. AI companies that build large language models (LLMs) and chatbots have created versions of their models adapted for government work, such as Anthropic's Claude Gov, designed to operate in more languages and in secure environments.
However, the official’s comments are the first indication that AI companies building LLMs, such as OpenAI and xAI, could train government-specific versions of their models directly with classified data. iFixit, who directs the Wadhwani AI Center at the Center for Strategic and International Studies and previously led AI policy efforts at Google and OpenAI, notes that training with classified data, as opposed to simply answering questions about it, would present new risks.
The most significant risk, he explains, is that the classified information used to train these models could reappear to anyone using the model. This could be a problem if various military branches, each with different classification levels and information needs, shared the same AI.
See also: US House of Representatives: Ban on the use of WhatsApp
However, Mehta argues that it is not that difficult to keep information restricted from the wider world: “If you set it up right, you have very little risk of this data appearing on the general internet or going back to OpenAI.”
