AI Model Escape, Then Target Tools to Help Themselves Improve

Goal-driven behavior, not malicious intent, is a key problem, according to security experts.