“LabelBox researchers were able to jailbreak top AI models, including those with strong reputations for safety, approximately 90% of the time using a new safety benchmark.”