This research reveals a critical security vulnerability in Multimodal Large Language Models (MLLMs) that allows attackers to manipulate model responses through simple negation techniques.

• Models frequently accept contradictions to their initially correct answers when faced with negation arguments
• Researchers developed GaslightingBench, a specialized benchmark to evaluate model robustness
• Even state-of-the-art MLLMs show significant performance drops when challenged with negation
• These findings highlight urgent security concerns for multimodal AI systems in high-stakes applications

For security professionals, this research emphasizes the need for more robust defenses against conversational manipulation in multimodal systems before deployment in critical environments.

Original Paper: Calling a Spade a Heart: Gaslighting Multimodal Large Language Models via Negation