This research explores how Multimodal Large Language Models can detect and reason about anomalies in images without requiring training on normal samples.

• Introduces a novel zero-shot approach that makes anomaly detection practical for data-restricted scenarios
• Demonstrates MLLMs' capability to identify visual abnormalities across various contexts including security applications
• Presents a new framework and benchmark for evaluating anomaly detection performance
• Shows promising results for security applications where detecting unusual patterns with minimal training is crucial

For security professionals, this research offers a path to implement anomaly detection systems in environments where collecting comprehensive training data is challenging or impossible.

Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models