Microsoft's new generative AI model for top-secret work with US intelligence agencies

New generative AI model for top-secret work / Illustartion / aidigitalx

It is designed to operate in an “air-gapped” or completely isolated environment, without any connection to the internet or cloud. This prevents any risk of hacking, data leaks, or unintended learning from external sources.

According to Bloomberg, Microsoft spent 18 months developing a secret, air-gapped AI model based on GPT-4 specifically for U.S. intelligence agencies like the CIA. This model is not connected to the internet for security reasons.

The model can answer questions, write code, read and analyze files, but cannot learn from new data to prevent sensitive information from entering it. It still needs to be tested and accredited by the agencies.

It runs on a specialized supercomputer in Iowa on a secure government network inaccessible to anyone outside approved US intelligence agencies.

The isolation prevents the model from inadvertently learning from user data, web data, or any other sources beyond its initial training. It operates on a “static cloud.”

U.S. intelligence agencies see great potential in using AI to process and analyze the massive amounts of data they collect more efficiently. This is viewed as a “strategic shift” in how they handle data.

The CIA specifically is hiring AI specialists and participated in launching the AIM Initiative in 2019 to apply AI to intelligence data processing and production.

A former CIA official wrote that current AI chatbots are innovative but not truly intelligent, but have potential for processing the variety of data intelligence agencies collect.

There appears to be a race among global intelligence agencies to be the first to effectively deploy generative AI for enhancing intelligence analysis while maintaining secrecy. Also to apply generative AI capabilities like ChatGPT to intelligence data sets, with the U.S. aiming to take the lead.