Scholarlink

The increasing reliance on Large Language Models (LLMs) in security-sensitive domains necessitates a thorough evaluation of their robustness against adversarial attacks. To address the existing gap in diverse, accurate, and up-to-date test datasets, we have created a novel dataset specifically designed to assess LLMs’ resistance to various prompt-based attack scenarios. This Security Evaluation dataset, SecEval, covers a range of security domains, including Physical Security, where LLMs might be leveraged to devise methods for compromising physical assets; Data Security, assessing the potential to undermine data integrity and availability; Application and Network Security, testing their capability to generate exploits and infiltrate networks; and specialized areas such as Endpoint Security, Identity and Access Security, and Operational Security. During the dataset creation, we employed a multi-phase quality control process, including initial filtering by GPT, manual review, testing with Llama 3.1, and iterative human refinement to ensure relevance and accuracy. We then conducted a detailed evaluation of state-of-the-art LLMs: Llama 3.1 and Gemma 2, focusing on their vulnerability to these attack types and their response efficiency. Our analysis differentiates between single-query attacks and sustained multi-query strategies, providing valuable insights into how the Llama 3.1 and Gemma 2 models adapt and maintains their security posture under continuous threats. The dataset is made publicly available at https://github.com/VeraaaCUI/SecEval-Dataset , offering a valuable tool for the research community to further investigate and evaluate the security of LLMs.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SecEval: A Security Evaluation Dataset for Large Language Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Huining Cui,
  • Wei Liu

摘要

The increasing reliance on Large Language Models (LLMs) in security-sensitive domains necessitates a thorough evaluation of their robustness against adversarial attacks. To address the existing gap in diverse, accurate, and up-to-date test datasets, we have created a novel dataset specifically designed to assess LLMs’ resistance to various prompt-based attack scenarios. This Security Evaluation dataset, SecEval, covers a range of security domains, including Physical Security, where LLMs might be leveraged to devise methods for compromising physical assets; Data Security, assessing the potential to undermine data integrity and availability; Application and Network Security, testing their capability to generate exploits and infiltrate networks; and specialized areas such as Endpoint Security, Identity and Access Security, and Operational Security. During the dataset creation, we employed a multi-phase quality control process, including initial filtering by GPT, manual review, testing with Llama 3.1, and iterative human refinement to ensure relevance and accuracy. We then conducted a detailed evaluation of state-of-the-art LLMs: Llama 3.1 and Gemma 2, focusing on their vulnerability to these attack types and their response efficiency. Our analysis differentiates between single-query attacks and sustained multi-query strategies, providing valuable insights into how the Llama 3.1 and Gemma 2 models adapt and maintains their security posture under continuous threats. The dataset is made publicly available at https://github.com/VeraaaCUI/SecEval-Dataset , offering a valuable tool for the research community to further investigate and evaluate the security of LLMs.