Overview
Severity: HIGH | Affected: Multi-modal LLMs | Category: research
A paper published by researchers at Stanford's AI Lab details a new jailbreak technique named 'Cognitive Dissonance Injection' (CDI). The method effectively bypasses safety filters on state-of-the-art multi-modal AI systems by providing conflicting inputs across different modalities. For example, an image of a benign object is submitted alongside a text prompt containing layered, harmful instructions subtly encoded within a seemingly innocuous request. The model's struggle to reconcile the contradictory signals from the image and text creates a temporary collapse in its safety alignment, allowing it to process and execute the harmful instructions. The research demonstrates a high success rate against several major commercial models, raising significant concerns about the robustness of current multi-modal safety architectures and prompting urgent reviews from developers.