AI safety group METR finds flaws in OpenAI’s internal monitoring system