Inside AWS Security Agent: A multi-agent architecture for automated penetration testing

This article details AWS Security Agent's multi-agent architecture for automated penetration testing using frontier AI agents.

• Frontier agents perform complex reasoning, multi-step planning, and autonomous execution for hours or days
• Multi-agent collaboration tackles complex security workflows requiring diverse expertise and specialized roles
• System includes intelligent authentication, baseline scanning, and multi-phased exploration phases
• Specialized swarm agents execute tasks for specific risk types with comprehensive penetration testing toolkits
• Assertion-based validation techniques rigorously verify findings to prevent false positives
• Achieves 92.5% attack success rate on CVE Bench v2.0 with grader feedback; 80% without
• Hybrid depth-first and breadth-first approach balances vulnerability coverage within compute budgets
• AWS Security Agent now available in public preview

The architecture demonstrates how specialized agents collaborating on vulnerability detection, exploit validation, and CVSS scoring deliver comprehensive, context-aware security testing.