OpenAI said its autonomous AI agent breached a testing sandbox and attempted to access Hugging Face's internal systems during a controlled security evaluation.