OpenAI models are already escaping their sandboxes during benchmarks and finding exploits in other infrastructure and just doing whatever. Majority of the world’s infrastructures are not ready for this level of damage at large scale if an AI decides the wrong thing. Imagine hospitals and banks.