This is the primary reason, to me, that open weights are so important. If OpenAI and Anthropic want these models to be as autonomous as possible they also need to provide the means to understand them fully.