New MemJack framework exploits natural images for VLM jailbreaks

Researchers have developed MemJack, a novel framework designed to test the safety of Vision-Language Models (VLMs) using natural images. This memory-augmented multi-agent system discovers reusable visual anchors within benign images to craft jailbreak attacks. MemJack has also led to the creation of MemJack-Bench, a dataset containing over 113,000 multimodal jailbreak trajectories for evaluating and improving VLM safety alignment. In tests, MemJack demonstrated significant effectiveness, achieving a 71.48% attack success rate against Qwen3-VL-Plus and highlighting substantial vulnerabilities in current safety-aligned VLMs. AI

IMPACT Highlights significant vulnerabilities in current VLM safety alignment, potentially driving new defensive strategies.

RANK_REASON The cluster contains an academic paper detailing a new method for attacking AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• COCO val2017
• Jian-Hao Chen
• MemJack
• MemJack-Bench
• Qwen3-VL-Plus
• Vision--Language Models

• safety
• paper
• model release

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