OpenAI cribbing from other researchers. We just have to assume OpenAI is actively adversarial in future. Accidental cyber intrusion is also well within model capability.