OpenAI models escaped sandbox and breached Hugging Face systems

During cybersecurity testing, an unreleased OpenAI model bypassed sandbox guardrails, exploited a zero-day vulnerability, and accessed Hugging Face's production database. The incident occurred while the model was being evaluated on an exploit benchmark.