It’s the first time a major large language model has operated fully separated from the internet, a senior executive at the US company said. Most AI models including OpenAI’s ChatGPT rely on cloud services to learn and infer patterns from data, but Microsoft wanted to deliver a truly secure system to the US intelligence community. 

Spy agencies around the world want generative AI to help them understand and analyze the growing amounts of classified information generated daily, but must balance turning to large language models with the risk that data could leak into the open — or get deliberately hacked. 

Microsoft has deployed the GPT4-based model and key elements that support it onto a cloud with an “air-gapped” environment that is isolated from the internet, said William Chappell, Microsoft’s chief technology officer for strategic missions and technology.

Intelligence officials have repeatedly made clear they hanker after the same generative AI tools that promise to revolutionize swaths of business and modernize economies. The CIA last year launched a ChatGPT-like service at unclassified levels — but the community wanted something capable of dealing with far more sensitive data. 

“There is a race to get generative AI onto intelligence data,” Sheetal Patel, assistant director of the CIA for the Transnational and Technology Mission Center, told delegates to a security conference at Vanderbilt University last month. The first country to use generative AI for their intelligence would win that race, she said. “And I want it to be us.”

Microsoft spent the past 18 months working on its system, including overhauling an existing AI supercomputer in Iowa. Chappell, an electrical engineer who previously worked on microsystems for the Defense Advanced Research Projects Agency, described the endeavor as a passion project and said his team wasn’t sure how to go about it when they started out in 2022.

“This is the first time we’ve ever had an isolated version – when isolated means it’s not connected to the internet – and it’s on a special network that’s only accessible by the US government,” Chappell told Bloomberg News ahead of an official announcement later on Tuesday. 

The GPT4 model placed in the cloud is static, meaning it can read files but not learn from them, or from the open internet. This way, Chappell said, the government can keep its model “clean,” and prevent secret information from getting absorbed into the platform. About 10,000 people would theoretically be able to access the AI, he said.

“You don’t want it to learn on the questions that you’re asking and then somehow reveal that information,” he said.

The service, which went live on Thursday, now needs to undergo testing and accreditation by the intelligence community. The Central Intelligence Agency and the Office of the Director of National Intelligence, which oversees America’s 18 intelligence organizations, did not immediately respond to requests for comment.

“It is now deployed, it’s live, it’s answering questions, it will write code as an example of the type of thing it’ll do,” said Chappell.