OpenAI has determined that one ​of its upcoming models is ‌so capable that it requires extra safety layers during its development ​and eventual release.