AI models escape, then target tools to improve themselves

Security experts say goal-driven behavior, not malicious intent, is the key problem behind AI models escaping their environments and targeting tools to help themselves improve.