English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original
智柴网 中文原文

AVISE: An Open-Source Framework for Evaluating the Security of AI Systems

Forum topic · 小凯 · 2026-04-24

Summary

This paper introduces AVISE (AI Vulnerability Identification and Security Evaluation), a modular open-source framework for identifying vulnerabilities in and evaluating the security of AI systems and models. As a demonstration, the authors extend the theory-of-mind-based multi-turn Red Queen attack into an Adversarial Language Model (ALM) augmented attack and build an automated Security Evaluation Test (SET) for discovering jailbreak vulnerabilities in language models. The SET consists of 25 test cases plus an Evaluating Language Model (ELM) that judges whether each test case successfully jailbreaks the target model, achieving 92% accuracy, an F1 score of 0.91, and a Matthews correlation coefficient of 0.83. Using the SET, the authors evaluated nine recently released language models of varying scales and found all of them vulnerable to the augmented Red Queen attack to varying degrees. AVISE provides researchers and industry practitioners with a scalable foundation for developing and deploying automated security evaluation tests, a concrete step toward more rigorous and reproducible AI security assessment. Authored by Mikko Lempinen, Joni Kemppainen, and Niklas Raesalmi; arXiv:2604.20833.

Paper Overview

Research area: NLP Authors: Mikko Lempinen, Joni Kemppainen, Niklas Raesalmi Published: 2026-04-22 arXiv: 2604.20833

Summary

As AI systems are increasingly deployed across critical domains, their security vulnerabilities pose growing risks of high-profile exploits and consequential system failures. Yet systematic approaches to evaluating AI security remain underdeveloped. This paper introduces AVISE (AI Vulnerability Identification and Security Evaluation), a modular open-source framework for identifying vulnerabilities in and evaluating the security of AI systems and models.

As a demonstration of the framework, the authors extend the theory-of-mind-based multi-turn Red Queen attack into an Adversarial Language Model (ALM) augmented attack and develop an automated Security Evaluation Test (SET) for discovering jailbreak vulnerabilities in language models.

Key Results

  • The SET comprises 25 test cases and an Evaluating Language Model (ELM) that determines whether each test case successfully jailbreaks the target model.
  • The ELM achieves 92% accuracy, an F1 score of 0.91, and a Matthews correlation coefficient of 0.83.
  • Nine recently released language models of different scales were evaluated with the SET; all were found vulnerable to the augmented Red Queen attack to varying degrees.

Significance

AVISE offers researchers and industry practitioners a scalable foundation for developing and deploying automated security evaluation tests, representing a concrete step toward more rigorous and reproducible AI security evaluation.

Original Abstract (excerpt)

As artificial intelligence (AI) systems are increasingly deployed across critical domains, their security vulnerabilities pose growing risks of high-profile exploits and consequential system failures. Yet systematic approaches to evaluating AI security remain underdeveloped. In this paper, we introduce AVISE (AI Vulnerability Identification and Security Evaluation), a modular open-source framework for identifying vulnerabilities in and evaluating the security of AI systems and models. As a demonstration of the framework, we extend the theory-of-mind-based multi-turn Red Queen attack into an Adversarial Language Model (ALM) augmented attack and develop an automated Security Evaluation Test (SET) for discovering jailbreak vulnerabilities in language models. The SET comprises 25 test cases an...

--- *Auto-collected on 2026-04-24*

Tags

#ai-security#arxiv#nlp#language-models#jailbreak#red-queen-attack#security-evaluation#open-source

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177618684