EventHallusion introduces a pioneering approach to diagnose and evaluate event hallucinations in video-based large language models, addressing a critical gap in multimodal AI security.

• Establishes the first benchmark specifically targeting event hallucinations in VideoLLMs
• Reveals concerning frequency of AI systems fabricating events that never occurred in videos
• Identifies key hallucination patterns through systematic testing methodologies
• Demonstrates how these hallucinations create security vulnerabilities in video analysis applications

This research is vital for security professionals as AI systems increasingly analyze surveillance footage, authenticate video evidence, and make critical decisions based on video content. Undetected hallucinations could lead to false security alerts, missed threats, or compromised video-based authentication systems.

EventHallusion: Diagnosing Event Hallucinations in Video LLMs