This research introduces a groundbreaking framework for detecting abnormal behaviors in Large Language Models by analyzing hidden state patterns, achieving over 95% detection accuracy.

• Identifies hallucinations, jailbreak attempts, and backdoor exploits through hidden state forensics
• Provides a practical solution for enhancing the security and reliability of LLM applications
• Addresses critical vulnerabilities being exploited by malicious actors in deployed systems
• Offers a new protective layer for organizations deploying LLMs in sensitive environments

This innovation is particularly valuable for security teams as it enables real-time monitoring of LLM outputs without requiring model retraining or architectural changes, significantly reducing security risks in production environments.

Exposing the Ghost in the Transformer: Abnormal Detection for Large Language Models via Hidden State Forensics