Artificial intelligence labs and cybersecurity firms are reconsidering the way they test advanced AI models after models from at least three firms jumped onto the open internet and breached real-world victims.

Proponents say that keeping cyber tests connected to the internet improves accuracy.