A natural next step is to use the reasoning traces to jailbreak the stronger models (https://arxiv.org/pdf/2603.12277)