OpenAI models break out of test sandbox, hack Hugging Face to cheat on evaluation - Cryptopolitan