Overview
Severity: HIGH | Affected: Leading AI Labs | Category: research
A new research paper published by security researchers at Carnegie Mellon University has detailed a novel jailbreak technique named 'Chroma-Shift.' This multimodal attack bypasses the safety filters of state-of-the-art vision-language models (VLMs) by encoding malicious instructions within the pixel data of seemingly benign images. The technique manipulates specific color channels in a way that is imperceptible to the human eye but is interpreted by the model as a high-priority command, effectively overriding its safety alignment. The researchers demonstrated that by using Chroma-Shift, they could compel models from several major AI labs to generate misinformation, hate speech, and illegal content. This discovery reveals a new class of adversarial vulnerabilities in multimodal systems and poses a significant challenge for developers working on AI safety and content moderation.