A safety-research prompt became a jailbreak that felled 9 of 23 models
A MATS researcher who built a prompt to generate synthetic transcripts found that a few hours of tweaking turned it into a cross-model universal jailbreak. Tested across 23 models, it hit 84 to 100 percent attack success on the nine most vulnerable, with only recent model updates resisting. The author shelved the planned open-source release. The lesson: even benign "safety" prompts sit a few edits away from weaponization.