Bartosz Cywiński comments on Censored LLMs as a Natural Testbed for Secret Knowledge Elicitation