Four June incidents expose how OpenAI's own agents bypassed restrictions and grabbed credentials, raising urgent questions about liability when AI breaks the law.