Securing large language models: A quantitative assurance framework approach

doi:10.1016/j.jisa.2025.104351

2026-01-21

Sander Stamnes Karlsen

Muhammad Mudassar Yamin

Ehtesham Hashmi *

Basel Katt

Mohib Ullah

DOI：10.1016/j.jisa.2025.104351

原文链接

查看原文

摘要

En 中文

Large Language Models (LLMs) are increasingly integrated into sensitive domains such as healthcare and autonomous systems, yet adoption is constrained by security risks that conventional assurance methods do not capture. Traditional software assurance techniques are inadequate for LLM-specific vulnerabilities, including prompt injection, insecure output handling, and training data poisoning. We introduce a quantitative security assurance framework for LLM applications that translates security requirements and vulnerabilities into measurable scores. The framework computes an Assurance Metric (AM) as AM = RM - VM, where VM is weighted using CVSS v4.0, and maps results to five security assurance levels, making security posture comparable, auditable, and actionable. Requirements span input/output validation, training data, development and deployment, access control, third-party services, and security procedures; vulnerability tests align with the OWASP Top 10 for LLMs (prompt injection, insecure output handling, training data poisoning, denial of service, sensitive information disclosure, overreliance, and model theft). Case study results show uncensored models (e.g., Llama2-uncensored) exhibit significantly higher exposure, especially to prompt injection and output-handling attacks-while censored and fine-tuned models attain higher assurance levels. Significance and impact: the framework provides transparent, quantitative scoring to compare systems, prioritize mitigations, and support evidence-based deployment and governance in high-takes environments, with continuous human oversight emphasized.

Keyword:

Large language models

Security assurance framework

Quantitative security assessment

LLM vulnerabilities

Security architecture