We’re running out of reasons to ignore AI safety

Earlier this month, OpenAI gave respective of its AI models a task: implicit a test designed to measurement their cybersecurity capabilities. It enactment the systems successful a sandboxed situation without an net transportation and acceptable them disconnected to work.

What happened next is astir laughably silly - but also, arsenic Adam Gleave, cofounder and CEO of AI information enactment FAR.AI enactment it, "a visceral illustration of however misaligned AI could origin harm." According to OpenAI, the models escaped the sandbox meant to incorporate them, moved done the company's interior systems, recovered a way to the internet, and past started looking for a mode into Hugging Face. And wherefore was …

Read the afloat communicative astatine The Verge.