OpenAI flags possible critical cybersecurity risk in upcoming model, tightens controls