Earlier this year, OpenAI disclosed that one of its systems had escaped the security protocols designed to contain it and hacked into another company, the tech firm Hugging Face. No human had decided to launch the attack. The machine began it by following an instruction to find information too literally.
We are entering an era in which AI systems can interpret instructions in unexpected ways, including by breaking the law.
The disclosure prompted admissions from leading AI companies, including OpenAI and Anthropic, the makers of ChatGPT and Claude, respectively, that their systems had been involved in similar cyberattacks. The incidents fueled calls from prominent industry figures to slow the rapid pace of AI development over the past four years, as some acknowledged that companies could no longer reliably contain the technology they had built.