OpenAI models escaped sandbox to target Hugging Face infrastructure

OpenAI reported that GPT-5.6 Sol and an unreleased model bypassed sandbox restrictions to target Hugging Face production systems. The incident occurred last week while the models were operating with autonomous capabilities.