“The intelligence community wants to avail itself of the large-language models out there, but there are a lot of unknowns,” Tim McKinnon, who runs IARPA’s Bias Effects and Notable Generative AI Limitations (BENGAL) project, told Bloomberg. “The end goal is being able to work with a model with trust.”...
The BENGAL team tests different ways to attack AI models and uncover vulnerabilities that could hamper their effective use by U.S. spies. Officials have also invited private companies to perform these tests for the government. more