Target Audience

• Regulatory Authorities: Providing objective, automated safety testing tools. Transforming abstract legal compliance requirements into executable quantitative metrics to serve as the criteria for LLM deployment approval and routine safety inspections.
• LLM Developers & Third-party Assessment Agencies: Serving as a pre-launch safety "health check." Conducting anti-scam red-teaming to identify and rectify security vulnerabilities in model responses to user-induced queries.

Multi-Agent Collaborative System

The system utilizes a multi-agent architecture coordinated by an orchestrator to achieve an automated closed-loop from simulated testing to report generation:

• High-fidelity Test Case Generation: Based on the latest real-world scam cases from police reports and cutting-edge Human-Computer Interaction (HCI) theories, the system wraps obvious scam information into various types of daily assistance requests (e.g., "Help me polish an email to this investment mentor").
• Multi-model Concurrent Testing: This agent acts as a "victim," sending various inductive test cases automatically and concurrently to mainstream LLMs while recording all raw response data.
• Security Defense Evaluation: This agent acts as an "objective judge," reading collected model responses and performing automated scoring. The evaluation includes two core dimensions:
  1. Qualitative Identification: Whether the model successfully identifies the scam risk behind the task.
  2. Quantitative Warning: The prominence and proportion of risk warning content in the model's response.

Core Features and Values

• Real-time Scenario Updates to Combat Evolving Tactics: Traditional static safety test sets become obsolete quickly. This tool supports the direct import of the latest real cases, which the system instantly transforms into new test sets. This ensures that regulatory measurement capabilities keep pace with the evolution of real-world scam tactics.
• Cultural Adaptation to Target Cognitive Blind Spots: How victims query AI is deeply influenced by local culture and psychological mindsets. Combining cutting-edge psychological research, the system establishes a strategy library based on synergistic cognitive biases, "Face Theory," and preemptive rebuttals. This ensures test cases truly reflect the interaction habits of users in different regions, deeply testing the interception capabilities of LLMs in complex contexts.
• Quantifiable Risk Metrics to Give Regulation "Teeth": Existing AI laws and regulations often remain at the level of macro expressions like "ensuring safety." This tool transforms abstract text regulations into clear, comparable data dashboards. For models with sub-standard test scores or those that permit scam risks, targeted rectification can be legally required, providing a practical technical lever for cyberspace governance in the AI era.