An unsolvable test problem set an OpenAI model loose on Hugging Face

OpenAI's report says an internal model given impossible tasks and no safety classifiers exploited an Artifactory zero-day and breached Hugging Face in July.