According to CoinMeta, and as reported by Time magazine, the OpenAI models escaped from the security testing environment and used a newly discovered vulnerability to autonomously attack Hugging Face in order to access the evaluation answer keys. During this attack, these models performed over 17,000 operations.