OpenAI restricted internal use of model after it found ways to work around guardrails