Microsoft has rolled out a GPT-4-based AI model tailored specifically for US intelligence agencies, as reported by Bloomberg. This groundbreaking move signifies Microsoft’s first deployment of a major language model in a secure environment, allowing spy agencies to analyze highly classified data without the risks associated with internet connectivity. Additionally, the model enables secure conversations with a chatbot akin to ChatGPT and Microsoft Copilot.
GPT-4, developed by OpenAI, is a large language model (LLM) capable of predicting likely tokens in a sequence. It has diverse applications, including crafting computer code and analyzing information. When configured as a chatbot, GPT-4 can power AI assistants that engage in human-like conversations. Microsoft obtained a license to use this technology through a deal with OpenAI, which stemmed from significant investments made by Microsoft in the organization.
The new AI service, which is currently unnamed publicly, meets the increasing demand among intelligence agencies for generative AI tools to process classified data securely. Unlike conventional cloud-based implementations like ChatGPT, this service operates on an isolated network and AI supercomputer modified over 18 months by Microsoft in Iowa. While it can read files provided by users, it cannot access the open internet, ensuring heightened security for sensitive information.
According to William Chappell, Microsoft’s chief technology officer for strategic missions and technology, the isolated version of GPT-4 is exclusively accessible by the US government on a dedicated network. The service was recently activated and is undergoing testing by relevant agencies within the intelligence community, with approximately 10,000 individuals currently having access.
Despite its potential, there are concerns regarding the accuracy of GPT-4 when analyzing critical data. Due to the inherent limitations of AI language models, such as the propensity to confabulate and provide inaccurate information, careful oversight and augmentation with external data sources are essential. Microsoft has not disclosed details regarding oversight, usage limitations, or auditability of the system, prompting further inquiries into its implementation.
