CREST-Search: Comprehensive Red-teaming for Evaluating Safety Threats in Large Language Models Powered by Web Search

2025-10-09

AI Summary

Web-search-enhanced large language models (Web-LLMs) face critical security risks—particularly vulnerabilities arising from adversarial prompts interacting with untrusted web content. Method: We propose CREST-Search, the first dedicated red-teaming framework for Web-LLMs. It introduces WebSearch-Harm, a curated dataset of harmful web search queries; leverages in-context learning to generate adversarial queries; incorporates an iterative feedback mechanism to refine attack efficacy; and fine-tunes a red-teaming agent to systematically bypass safety filters. Contribution/Results: Experiments demonstrate that CREST-Search significantly improves vulnerability detection rates across mainstream Web-LLMs, effectively exposing their defensive weaknesses in realistic search-augmented interactions. The framework underscores the necessity of search-specific security evaluation methodologies and motivates the development of tailored defense mechanisms for web-enhanced LLMs.

Technology Category

• Computer Vision: Adversarial Attacks & Robustness
• Natural Language Processing: Safety and Robustness
• Search and Optimization: Adversarial Search

Application Category

• Search and Retrieval-Augmented AI: Web search models and ranking
• Responsible Web: Mitigating online harms, including harassment, spam, and fake reviews
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Large Language Models (LLMs) excel at tasks such as dialogue, summarization, and question answering, yet they struggle to adapt to specialized domains and evolving facts. To overcome this, web search has been integrated into LLMs, allowing real-time access to online content. However, this connection magnifies safety risks, as adversarial prompts combined with untrusted sources can cause severe vulnerabilities. We investigate red teaming for LLMs with web search and present CREST-Search, a framework that systematically exposes risks in such systems. Unlike existing methods for standalone LLMs, CREST-Search addresses the complex workflow of search-enabled models by generating adversarial queries with in-context learning and refining them through iterative feedback. We further construct WebSearch-Harm, a search-specific dataset to fine-tune LLMs into efficient red-teaming agents. Experiments show that CREST-Search effectively bypasses safety filters and reveals vulnerabilities in modern web-augmented LLMs, underscoring the need for specialized defenses to ensure trustworthy deployment.

Problem

• Evaluating safety threats in web-augmented large language models
• Systematically exposing risks in search-enabled LLM workflows
• Addressing vulnerabilities from adversarial queries with untrusted sources

Innovation

• Generates adversarial queries using in-context learning
• Refines queries through iterative feedback mechanisms
• Fine-tunes LLMs with WebSearch-Harm dataset