This "adverserial" style of evaluating an LLM is interesting, but not the best evaluation. The best evaluation for how good an LLM is is if you give your best prompt instead of antagonizing the chatbot. That's the best way how we find out what its maximum capabilities are.