Red VS Blue Agents — Security skill for Claude Code

Red Team vs Blue Team: Multi-Agent Code Security Duel

Two Claude-powered agents compete round by round to harden a piece of code:

• 🔴 Red agent (ethical pentester) audits the code and returns a structured vulnerability report: severity, OWASP category, location, attack vector and recommendation.
• 🔵 Blue agent (defender) receives the report and patches the code, explaining every change.
• ⚖️ Orchestrator repeats Red → Blue, scores each round by severity and picks a winner:
  • Blue wins when Red can no longer find any CRITICAL or HIGH issues.
  • Red wins if the round limit is reached with blocking issues still open.

This is **defensive** security. The red agent audits *your own* code so the blue agent can fix it. It writes reports and fixes, never working exploits or malware.

Two implementations, one idea

The project implements the same duel twice to compare two ways of building agents:

	duelo_seguridad.py	duelo_seguridad_sdk.py
Approach	Raw Anthropic Python SDK (Messages API)	Claude Agent SDK
How agents work	Pure text in, JSON out	Agents use tools: read files, grep, run Bash and bandit (SAST), edit files in place
Blue's output	Full patched code inside JSON	Edits the files directly, can run tests
Scope	One file or the built-in demo	A file or a whole folder (worked on a copy in _work_duelo/)
Extras	Saves the hardened code to a file	Final cleanup pass for remaining MEDIUM/LOW issues, JSON reports per round

Building the raw version first made the agent loop explicit: prompting, structured output, parsing and orchestration are all hand-written. The Agent SDK version shows how much changes once agents can act on a real codebase.

Example run

On the built-in demo (a small Flask app with SQL injection, path traversal, a hardcoded secret and debug mode on), the red agent's score dropped each round until the blue agent won:

| Round | Red's findings | Score | |--