Strict safety filters on closed AI models do more harm than good to cybersecurity