Experts tried to get AI to create malicious security threats – but what it did next was a surprise even to them

Experiments find LLMs can create harmful scripts, although real-world reliability failures prevent them from enabling fully autonomous cyberattacks today.