you know it says quite a lot about all these big frontier models, that they pull this sort of shit: - they are asked to do a thing (like participate in a ctf) - they try, and often they realize they cant do the thing - their base training says 'satisfy the user' - that results in "must satisfy user at all costs" behavior - cant solve the puzzle? cheat! - in this case, 'cheating' involved hacking the place hosting the contest files to steal the answers.