Comment by hellojesus

Maybe their methodology worked at the start but has since stopped working. I assume model outputs are passed through another model that classifies a prompt as a successful jailbreak so that guardrails can be enhanced.