“During a test, OpenAI intentionally reduced the guardrails on two of its advanced AI models within a sandboxed, closed environment.”