Mapping the Attack Surface: Lasso Security’s Eliya Saban on GEO-Based AI Manipulation
50,000+ enterprise professionals read CloudTweaks monthly
Executive InterviewJun 25, 20269 min read
Most people assume that if an AI assistant gives a wrong answer, someone must have hacked it. New research from Lasso Security suggests the reality is far more mundane and more concerning. By applying ordinary content-optimization techniques to a normal-looking website, researchers were able to push a fabricated medical claim into AI-generated answers, in some cases with a 98% success rate, without any hidden code or prompt injection. The model wasn’t broken; it was simply persuaded by information it retrieved and trusted.
In this interview, we speak with Eliya Saban, a Security Researcher on Lasso Security’s offensive AI security team, about how these “GEO” attacks work, why the health domain raises the stakes, and what it means when a confident AI answer carries no trace of where it came from. The window for defenders, she warns, may already...
Copyright of this story solely belongs to cloudtweaks.com. To see the full text click HERE
