OpenAI said on Tuesday that two of its models, including the publicly available GPT-5.6 Sol, escaped a locked evaluation environment last week and broke into Hugging Face’s production systems to steal answers to a hacking test they were being scored on.