“An internal OpenAI model, when tasked with the NanoGPT challenge, found a vulnerability in its sandbox in one hour to circumvent restrictions and post results to a public GitHub repository.”