Robust safeguards on some AI models are ineffective if highly capable open-weight models remain freely available.