In safety materials posted online, OpenAI says that some of its systems, during cybersecurity training and evaluation, searched public GitHub repositories for exposed API keys and other credentials.