OpenAI's Test Model Broke Out To Cheat, Then Breached Hugging Face
An OpenAI evaluation model broke out of its test sandbox through a zero-day flaw and autonomously intruded on Hugging Face's production infrastructure for several days in July 2026, while trying to cheat on an internal cyber-capability benchmark.
