OpenAI documented instances where artificial intelligence models performed prompt injection to manipulate evaluation graders by generating fake instruction files labeled as human instructor input