AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Efficient LLM-Jailbreaking via Multimodal-LLM Jailbreak

Haoxuan Ji, Zheng Lin, Zhenxing Niu, Xinbo Gao, Gang Hua

AAAI 2026 · Large language models › Language, safety and interpretability · approach: Jailbreak attacks and defenses · 5 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
jailbreak LLM via MLLM jailbreak embedding instead of GCG token searchLLaMA2-Chat-7B · AdvBench-M testavg ASR (%): 74.7 -> 84.7Improved
use MLLM-jailbreak approach instead of AdvAttackInternLM2.5-7B · AdvBench-M testavg ASR (%): 58.0 -> 84.6Improved
use MLLM approach instead of GCG for runtimeLLaMA2-Chat-7B · per-class runtimehours: 11.2h -> 0.37hImproved
use image-text semantic matching for initial image instead of random (ablation)MLLM jailbreak · AdvBench-M Dtrain (MiniGPT4)ASR (%): 73.32 -> 83.70Improved
use image-text semantic matching instead of ranking-based init (ablation)MLLM jailbreak · AdvBench-M Dtrain (MiniGPT4)ASR (%): 75.18 -> 83.70Improved
optimize image via visual module instead of directly optimizing soft embeddings (ablation)embedding-based jailbreak · token conversioncosine similarity: 0.17 -> 0.63Improved
compare ours with AdvAttack on validation splitInternLM2.5-7B · HarmBench VALASR (%): 68.7 -> 59.0Hurt
transfer InternVL2.0-derived suffixes to proprietary modelsblack-box targets · AdvBench-Mavg ASR (%): 77.1 (Mistral) vs 20.1 (DouBao)Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy