Source: Rapid7 Cybersecurity Blog
Author: Brian Bartholomew
URL: https://www.rapid7.com/blog/post/so-red-teaming-offensive-methodology-multi-agent-ai-architecture
https://www.rapid7.com/blog/post/so-red-teaming-offensive-methodology-multi-agent-ai-architecture
ONE SENTENCE SUMMARY:
Rapid7 built a production multi-agent red-teaming system using frontier models to automate mechanics, keep humans in control, and improve AI defense.
MAIN POINTS:
- Attackers use AI to accelerate recon, vuln discovery, and scalable social engineering.
- Rapid7 formalized pentest workflow into a production multi-agent system, not a prototype.
- Project Glasswing provided early access to Claude Mythos for proactive security research.
- Frontier model plus structured architecture improved vulnerability analysis and exploit chaining quality.
- Goal: automate repeatable tasks while reserving critical judgement decisions for humans.
- Orchestrator coordinates specialists; routing separated from execution for auditability and control.
- Engagement methodology was reverse-engineered from real tester task lists into orchestration logic.
- Scope decomposition prevents shallow analysis by giving each component full context and attention.
- Feedback-triggered re-entry replaces linear pipelines, reflecting real pentest discovery loops.
- Tiered guardrails enforce scope, classify actions, and require approval for risky dynamic tests.
TAKEAWAYS:
- Institutional methodology, not the LLM itself, most strongly determines offensive agent effectiveness.
- Orchestration-first designs improve predictability, controllability, and forensic traceability in sensitive environments.
- Chunking targets enables depth, parallelism, and measurable coverage across complex applications.
- Replacing non-reasoning steps with scripts/MCP services cuts token costs and boosts practicality.
- Building offensive agents sharpens defensive insight into prompt injection, trust boundaries, and guardrail bypasses.