DEVOURED
From safety research prompt to cross-model universal jailbreak
A new universal jailbreak template successfully bypassed safety filters on 14 of 23 tested models with up to 100% success rates.
What: The attack utilizes a synthetic transcript generation prompt, exploiting consistent vulnerabilities across disparate LLM architectures.
Why it matters: The persistence of these jailbreaks across different companies suggests that current safety fine-tuning is brittle and potentially insufficient against standardized adversarial templates.
Decoder
- Universal jailbreak: An attack string that can bypass safety filters on a wide variety of different models, rather than just one specific instance.
Original article
A MATS researcher found that a synthetic transcript generation prompt could be turned into a universal jailbreak template that hit 84-100% attack success on the nine most vulnerable of 23 models tested, with only recent Anthropic models and Meta Muse Spark 1.1 never fully broken.
Explore all 57 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi