Quantifying the Irreducible: A Systematic Survey of Adversarial Jailbreak Vectors, Red Team Methodologies, and Enterprise Threat Exposure in Large Language Model Deployments

Key Points

• This research aims to systematically survey adversarial jailbreak attacks on large language models and propose a risk quantification framework.
• Reviewed 65 peer-reviewed publications from 2022 to 2025.
• Introduced the Enterprise Threat Exposure Model (ETEM) to quantify risks.
• Classified 23 distinct attack methodologies using a five-dimensional taxonomy.
• Conducted cross-model vulnerability analysis across six model families.
• Outlined a five-layer defense-in-depth architecture and identified seven open research challenges.
• Identified key adversarial jailbreak vectors affecting large language models.
• Developed indices like the Adversarial Penetration Index (API) to assess threats.
• Highlighted emergent threats from new architectural designs and exploitation techniques.
• Created a comprehensive classification of attack methodologies.
• Provided a framework for improved defense measures against identified vulnerabilities.