Mirage: Targeted Jailbreaks Against Post-Generation Safety Checkers in Multimodal Diffusion Models

Summary

Researchers present Mirage, a multimodal jailbreak attack that bypasses post-generation safety checkers in diffusion pipelines. Mirage hides an adversarial target, either a reference NSFW image or a malicious text description, inside an imperceptible perturbation on a benign input image, exploiting the shared representation space of multimodal encoders. The authors report targeted jailbreak success against Stable Diffusion 1.5, Stable Diffusion XL, and Kandinsky 2.2 using limited black-box queries.

Topics

• Prompt injection and jailbreaks

Related items

• InfoLawmakers probe growing use of Chinese AI models in U.S. companiesSame vendor · CNBC Technology
• InfoThe Atlantic created a searchable database of the music used to train AISame vendor · The Verge (AI)
• MediumMalicious JetBrains Plugins Steal AI API Keys as Chrome Extensions Capture Chatbot ChatsSame vendor · The Hacker News
• InfoWhite House memo claims mass AI theft by Chinese firmsSame vendor · BBC Technology
• InfoAs the US Midterms Approach, AI Is Going to Emerge as a Key Issue Concerning VotersSame vendor · Schneier on Security