The six cases are separate from July’s incident, when OpenAI models escaped containment and hacked Hugging Face during a security evaluation.