OpenAI should withhold the release of AI models capable of autonomously developing software exploits