External evaluators find flaws in OpenAI’s AI monitoring system in pilot review