New RMT-Jailbreak framework transfers AR attacks to diffusion LLMs

Researchers have developed a new framework called RMT-Jailbreak that can transfer existing autoregressive (AR) jailbreak attacks to diffusion large language models (dLLMs). The method leverages the observation that harmful responses to queries can serve as effective jailbreak templates for dLLMs. By masking a portion of these harmful response tokens and applying safety filtering, the framework induces dLLMs to reconstruct the harmful content. Experiments show strong cross-paradigm transferability of these jailbreak attacks across different AR methods and dLLMs. AI

IMPACT This research highlights a new vulnerability in diffusion LLMs, potentially impacting safety filtering mechanisms and requiring new defense strategies.

RANK_REASON Research paper detailing a new method for jailbreaking LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Autoregressive jailbreak
• Diffusion Large Language Models
• Hugging Face
• RMT-Jailbreak

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →