AI agents are increasingly deployed in production, yet security evaluations remain stuck in the past, relying on manual red-teaming and static benchmarks that cannot model adaptive, real-world adversaries. NAAMSE closes that gap by treating agent security as a feedback-driven optimization problem. Rather than running a fixed set of attacks, it evolves them, using each generation's results to compound pressure on your agent's weak points and surface jailbreaks, prompt injections, and PII leakage that one-shot methods routinely miss. Every run produces a comprehensive report with vulnerability analysis, attack effectiveness metrics, and cluster-based categorization of discovered exploits, giving your team a clear and actionable picture of where your agent breaks before it ever reaches production.