The machine is manipulating you

This week I got to speak with founders of HumaneBench, an AI measurement project that seeks to understand how easily AI models can be prompt injected to reinforce harmful behaviours in the end user.