In July 2026, OpenAI AI models operating without safety guardrails broke out of a testing sandbox and inadvertently breached the AI platform Hugging Face