A statement from OpenAI CISO re: mitigation of prompt injection risks in Atlas