LiaisonAgent: An Multi-Agent Framework for Autonomous Risk Investigation and Governance
Organizations: Shenzhen Institute of Advanced Technology, CAS · Sangfor Technologies Inc. · College of Management Science · Chengdu University of Technology ·
Abstract
The rapid evolution of sophisticated cyberattacks has strained modern Security Operations Centers (SOC), which traditionally rely on rule-based or signature-driven detection systems. These legacy frameworks often generate high volumes of technical alerts that lack organizational context, leading to analyst fatigue and delayed incident responses. This paper presents LiaisonAgent, an autonomous multi-agent system designed to bridge the gap between technical risk detection and business-level risk governance. Built upon the QWQ-32B large reasoning model, LiaisonAgent integrates specialized sub-agents, including human-computer interaction agents, comprehensive judgment agents, and automated disposal agents-to execute end-to-end investigation workflows. The system leverages a hybrid planning architecture that combines deterministic workflows for compliance with autonomous reasoning based on the ReAct paradigm to handle ambiguous operational scenarios. Experimental evaluations across diverse security contexts, such as large-scale data exfiltration and unauthorized account borrowing, achieve an end-to-end tool-calling success rate of 97.8% and a risk judgment accuracy of 95%. Furthermore, the system exhibits significant resilience against out-of-distribution noise and adversarial prompt injections, while achieving a 92.7% reduction in manual investigation overhead.
Figures & tables
| Specific Tool | Functional Summary | Logic / Trigger |
| invest_ask_admin | Requests authorization from a security admin. | Triggered by high-severity technical alerts. |
| invest_notify_admin | Pushes unidirectional updates to administrators. | Used for status logging and final reporting. |
| invest_ask_user | Conducts multi-round Q&A with the user. | Initiated after administrator approval. |
| invest_ask_manager | Confirms business legitimacy with a supervisor. | Triggered if users’ feedback is suspicious. |
| invest_judge | Performs final risk qualification. | Executes after all feedback is collected. |
| invest_notify_user | Sends final status, security notification etc. | Triggered upon task completion. |
| Alert Category | Samples | OSR | FSR |
| Large Files Download | 4000 | 93.0% | 94.0% |
| Account Borrowing | 4000 | 99.0% | 99.7% |
| Other Anomalies | 8000 | 97.0% | 98.8% |
| Weighted Average | - | 96.5% | 97.8% |
| Agent Role | Score |
| Investigator for User Inquiry | 93% |
| Investigator for supervisor verification | 99% |
| Comprehensive Risk Judgment | 98% |
| Risk Disposal | 98% |
| LLM Model | Tool Calling | Planning Accuracy | Logical Reasoning | Investigation Dialogue |
| Qwen2.5-32B | 71.2% | 95.6% | 82.5% | 82.1% |
| Qwen3-14B | 72.3% | 90.8% | 80.9% | 79.2% |
| Qwen3-32B | 88.5% | 96.3% | 85.7% | 87.5% |
| Qwen3-30B-A3B | 90.1% | 94.5% | 81.6% | 85.7% |
| QwQ-32B | 96.5% | 97.8% | 89.6% | 87.3% |
| Answer Type | Sample Size | Defense Success Rate |
| Irrelevant Topics Answer | 50 | 95% |
| Gibberish & Semantic Voids | 50 | 94% |
| Emotion Appeals | 50 | 98% |
| Instruction Injection | 100 | 95% |
| White-Box Tool Spoof | 100 | 100% |
Autonomous Cyber Defense
Autonomous cyber defense uses AI agents to detect, investigate, and contain attacks while balancing operational safety, latency, and human oversight.
See all 72 papers