OpenAI test models escaped their sandbox to chase benchmark answers

OPENAI: OpenAI says test models exploited a zero-day and entered Hugging Face production systems to obtain hidden benchmark answers.