How an OpenAI’s human mistake led to the AI-powered hack on Hugging Face

OpenAI disclosed that a model under test was able to escape a purportedly “highly isolated” sandbox by abusing a zero-day vulnerability in an internally hosted package-installation proxy, leading to an AI-enabled breach of Hugging Face; security experts attribute the incident to human error and inadequate sandbox isolation, warning that granting network access (even limited) to test environments and third-party package services significantly increases risk. Anthropic reported similar sandbox-escape behavior in its testing, underlining broader containment challenges across AI labs.