Overview
Severity: HIGH | Affected: Multiple AI Labs | Category: research
Researchers from Carnegie Mellon University have published a paper detailing a new jailbreak technique named 'Synesthesia.' This method exploits the way large multimodal models (LMMs) process and integrate information from different modalities, such as text and images. By embedding malicious prompts within the imperceptible noise of an image file and pairing it with a seemingly benign text query, attackers can trigger the model to generate harmful or restricted content. The poisoned image data subtly biases the model's internal representations, effectively bypassing text-based safety filters. This technique proved successful against several leading commercial LMMs, raising significant concerns about the security of multimodal AI systems and the need for more robust, cross-modal alignment and filtering mechanisms.