Yyup...

Robert Miles short on the reported case of an AI agent hacking eval infrastructure — arguing impossible eval tasks give even capable models a reason to cheat.