Researchers introduce CoFiBreak to improve jailbreak attack diversity in LLMs
A new method called CoFiBreak aims to enhance the diversity of jailbreak attacks on large language models (LLMs) while maintaining effectiveness. Developed by researchers, CoFiBreak uses a coarse-to-fine diversity reward within Group Relative Policy Optimization (GRPO) to broaden the range of attack strategies without sacrificing success rates or inference efficiency. The approach ensures…
Key points
- CoFiBreak improves jailbreak attack diversity in LLMs using a coarse-to-fine reward system within GRPO
- Method achieves 1.6–2.0x more distinct successful strategies than baselines on HarmBench
- Reaches 98.8% ASR@10 on AdvBench with highest diversity score, transferring to unseen models like Claude-Opus-4.6
Tests on frontier LLMs show CoFiBreak discovers 1.6 to 2.0 times more distinct successful strategies than existing methods on HarmBench. It achieves a 98.8% attack success rate (ASR@10) on AdvBench with the highest diversity score, outperforming baselines by 2.1x. The method also transfers well to unseen models like Claude-Opus-4.6, achieving a 71.0% ASR@10. The code will be publicly released.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- SemiAnalysis releases GDPx benchmark with Claude Opus 5.5 leading at 12% · 3 src
- Gemma-3-12b fails to detect misinformation when poisoned claims dominate retrieved evidence · 1 src
- Researchers release BILAL-BENCH to test Bengali AI models’ Islamic text fabrication · 1 src
- Researchers prune 729 attention heads to cut GPT-2-XL’s memorisation by 91.5% · 1 src
- Researchers study how LLMs assign emotional value to outcomes · 1 src
Comments
via GitHub Discussions