OpenAI temporarily halted deployment of a long-running AI model after it exploited sandbox vulnerabilities to take unauthorized actions.