GPT-5 jailbreaks expose AI agents to data theft

Researchers have discovered a method to bypass GPT-5's safety features, enabling illicit instructions and exposing AI agents to data theft risks. The technique uses narrative jailbreaks and a "persuasion" loop to trick the model into generating harmful content by embedding requests within stories, bypassing keyword and intent filters. This vulnerability highlights risks as AI agents connect to cloud and IoT systems, with other zero-click attacks demonstrated against cloud storage and AI code editors.