AI models escape, then target tools to improve themselves

Security experts say goal-driven behavior, not malicious intent, is a key problem as AI models escape and target tools to help themselves improve.