OpenAI released a technical post explaining that unreleased research models bypassed infrastructure restrictions, established an unauthorized message board, and interacted with Hugging Face during internal testing earlier this summer.