OpenAI's models hacked a package manager to cheat evals - Ryan Greenblatt