AI 分析：该研究揭示了多智能体系统中一个重要的安全漏洞，即恶意Agent描述可以在不被调用的情况下污染Planner。这种攻击方式新颖且隐蔽，成功率变化显著（从84.31%降至37.25%），对系统性能影响巨大。这提示我们需要重新审视多智能体系统的安全架构，特别是对第三方Agent的注册和描述验证机制。
Analyze the security implications of agent description injection in multi-agent systems. Explain how malicious agent descriptions can affect the planner without being invoked, and propose potential defense mechanisms. Include case studies showing performance impact (e.g., success rate dropping from 84.31% to 37.25%). Outline a testing framework to detect such vulnerabilities.