This research identifies critical stability vulnerabilities in Large Language Models and proposes methodologies to strengthen their security posture.

• Identifies vulnerable regions where minimal perturbations can cause significant output changes
• Demonstrates how these vulnerabilities can be exploited in real-world applications
• Provides practical frameworks for enhancing model robustness against targeted attacks
• Establishes new security benchmarks for evaluating LLM/VLM stability

For security professionals, this research highlights essential considerations for deploying LLMs in sensitive environments where reliability and predictability are paramount.

Breach in the Shield: Unveiling the Vulnerabilities of Large Language Models