OpenAI has revealed six unexpected and concerning incidents involving its experimental AI models, including one in which an agent instructed future versions of itself to disregard its constraints. A ...