HalluDomainGuard Agent: A Domain Security Audit System for Large Language Model Responses

HalluDomainGuard Agent is a local audit system for AI security research scenarios, focusing on resolving security risks like hallucinated domains, brand impersonation, and phishing inducement in large language model (LLM) responses. Its core concept is 'evidence chain protection', which provides traceable and verifiable security audit conclusions through multi-layer network evidence collection and a deterministic rule engine, without relying on LLM's self-judgment of security.

With the widespread application of LLMs, users often use them as search engines to query information like official URLs. However, LLM responses may contain fake, misleading, or malicious domains (e.g., made-up official-like URLs, recommended hijacked domains), leading to user account theft, information leakage, or property loss. HalluDomainGuard Agent is precisely designed as a localized AI security audit system to address this challenge.

HalluDomainGuard Agent adopts a 9-stage pipeline architecture: 1. Domain Extraction (parse responses to extract domains and context); 2. Network Evidence Collection (parallel DNS resolution, TLS certificate verification, HTTP redirect tracking, RDAP query); 3. Threat Intelligence & Popularity Assessment (integrate intelligence sources like VirusTotal, Cloudflare popularity ranking); 4. Risk Rule Scoring (7 hard access control rules + 12 scoring rules); 5-9. Expert Model & Decision Fusion (optional GPT-5.5 semantic analysis, integrate rule-based and expert conclusions).

• Official Domain Truth Database: Contains 15 sources (31k entities, 35k domains) covering finance, government, and other fields. It includes offline sources (local seed database, banking and insurance systems, etc.) and online update sources (CISA, FDIC, etc.), stored in SQLite, and supports brand similarity detection.
• SSRF Protection: Blocks localhost, private network addresses, does not execute JS or submit forms, and sets timeout limits to ensure tool safety.

• Frontend Interface: Built with native HTML/CSS/JS, including a detection console, evidence visualization panel, expert correction comparison, and historical report management (export JSON/PDF).
• Ablation Experiments: 6 schemes (full chain, rule-only, etc.) support security research.
• Deployment: Requires only Python 3.11+ and PyMuPDF; optional external APIs (auto-degrade without keys); simple startup command and convenient API calls.

HalluDomainGuard Agent transforms the 'AI hallucination' problem into an observable and auditable process, providing a 'trust but verify' architectural approach and an engineering paradigm for AI security audit tool design. In the future, it can be extended to multi-modal scenarios (image QR codes, video URLs, etc.), and its modular architecture supports further expansion.