Diagnose Before You Refine: Multi-Dimensional Judge Feedback for Query-Efficient Jailbreak Attacks
A new paper reports a query-efficient jailbreak method with 89.8% HarmBench success in 10 queries.
An arXiv cs.CR paper introduces Jev-Guided Strategy Bandit, a query-based jailbreak framework that uses multi-dimensional probabilistic judge feedback to refine adversarial prompts. The authors evaluate it on two harmful-request datasets across five open-source and three closed-source LLMs. With a budget of 10 target-model queries, they report an average HarmBench attack success rate of 89.8% across ten open-source settings, above compared automated attacks. Defense tests cited limitations in existing filters and discussed Jev-based output filtering.
- JSB uses multi-dimensional judge scores to guide prompt refinement.
- With 10 target queries, reported average HarmBench success was 89.8%.
- Evaluation covered five open-source and three closed-source LLMs.
- Authors report filter weaknesses and discuss Jev-based output filtering.
From the original article
1 min read at arxiv.orgAutomated jailbreak attacks rely on evaluation feedback to optimize adversarial prompts, but existing methods often lack fine-grained diagnostic information. We propose Jev-Guided Strategy Bandit (JSB), a query-based jailbreak framework that leverages multi-dimensional probabilistic feedback to guide prompt optimization.
By Yixin Tan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.