Claude Code Security Agents: Automated Red Team + Blue Team

50+ vulnerabilities found and fixed using two Markdown files that turn Claude Code into offensive and defensive security specialists.

Drop two .md files into your project. Get automated pentesting and remediation — no frameworks, no infra, no security background required.

🇧🇷 Leia em Português

---

Table of Contents

1. The Problem
2. The Solution
3. How It Works
4. Setup (2 minutes)
5. Red Team — Offensive Agent
6. Blue Team — Defensive Agent
7. Running the Agents
8. Real Vulnerabilities Found
9. Customizing for Your Stack
10. FAQ

---

The Problem

You're building fast. Shipping features. But security keeps getting pushed to "later."

Meanwhile, your codebase quietly accumulates vulnerabilities: unsanitized HTML rendering, public API endpoints, admin routes without protection, webhook handlers that don't verify signatures. You won't find them in a code review — they hide in the interaction between components, in the assumptions you made at 2 AM, in the third-party integration you copy-pasted from the docs.

Hiring a pentester costs thousands. Running SAST tools gives you 200 generic warnings, most of them false positives. And you still have features to ship.

---

The Solution

Two specialized agents for Claude Code — each is a single Markdown file (~120 lines):

Agent	Role	Mode	What it does
Red Team	Attacker	Read-only	Scans your codebase for vulnerabilities. Thinks like a pentester. Generates a structured report with severity, attack scenarios, and fix recommendations
Blue Team	Defender	Edit mode	Takes the Red Team report and fixes every vulnerability. Writes code, adds tests, commits changes

They work as a feedback loop:

Red Team scans → finds 14 vulnerabilities → generates report
                                                    ↓
Blue Team reads report → fixes 12/14 → commits fixes
                                                    ↓
Red Team re-scans → finds 3 new issues → generates report
                                                    ↓
Blue Team fixes → hardening complete

Each cycle takes minutes, not days.

---

How It Works

Claude Code supports custom agents — Markdown files in .claude/agents/ that give Claude a specialized persona, methodology, and constrained tool access.

The Red Team agent gets read-only tools (Read, Grep, Glob, npm audit). It can analyze everything but change nothing.

The Blue Team agent gets edit tools (Read, Write, Edit, git). It can modify code, install packages, run tests, and commit.

Both agents inherit your project context from CLAUDE.md, so they understand your stack, architecture, and constraints without any extra configuration.

---

Setup (2 minutes)

1. Create the agents directory

mkdir -p .claude/agents

2. Copy the agent files

# Option A: Clone this repo and copy
git clone https://github.com/lucasrosati/claude-security-agents.git /tmp/security-agents
cp /tmp/security-agents/agents/red-team-scanner.md .claude/agents/
cp /tmp/security-agents/agents/blue-team-defender.md .claude/agents/

# Option B: Download directly
curl -o .claude/agents/red-team-scanner.md https://raw.githubusercontent.com/lucasrosati/claude-security-agents/main/agents/red-team-scanner.md
curl -o .claude/agents/blue-team-defender.md https://raw.githubusercontent.com/lucasrosati/claude-security-agents/main/agents/blue-team-defender.md

3. Run your first audit

Open Claude Code in your project and type:

Run the red-team-scanner agent on this codebase

That's it. No config files, no API keys, no Docker containers.

---

Red Team — Offensive Agent

File: agents/red-team-scanner.md

Model: Sonnet (fast, cost-effective for scanning) Permission mode: plan (read-only — cannot modify your code)

What it analyzes

Category	Examples
Authentication & Authorization	Broken auth, weak JWT validation, missing rate limiting, IDOR, privilege escalation
API Security	Public endpoints, missing object-level auth, mass assignment, enumeration
Input Validation & Injection	SQL/NoSQL injection, XSS (stored/reflected/DOM), SSRF, path traversal
File Uploads	MIME spoofing, missing magic byte validation, public storage exposure
Data Security & Compliance	Hardcoded secrets, PII in logs, insecure token storage, LGPD/HIPAA gaps
Infrastructure	CORS misconfiguration, missing security headers (CSP, HSTS), container issues
Admin Panel	Unprotected admin routes, hidden endpoints, session hijacking
Payment & Billing	Missing webhook verification, price manipulation, subscription bypass
Supabase-Specific	Broken RLS policies, service_role key exposure, cross-tenant access

Output format

For each vulnerability found:

[VULN-001] XSS via unsanitized AI content rendering
Severity: HIGH
Location: src/pages/StudySession.tsx:267
Description: AI-generated HTML rendered via dangerouslySetInnerHTML without sanitization
Attack scenario: Attacker injects <script> via manipulated study content → steals session tokens
Impact: Account takeover, data exfiltration
Fix: Sanitize with DOMPurify before rendering

---

Blue Team — Defensive Agent

File: agents/blue-team-defender.md

Model: Sonnet Permission mode: acceptEdits (can modify code with your approval)

What it fixes

Category	Actions
Authentication	Rate limiting, JWT hardening, session invalidation, role validation
API Security	Auth middleware, input validation (Zod/Joi), object-level authorization
Injection Prevention	DOMPurify for XSS, parameterized queries, path traversal checks, SSRF whitelists
File Uploads	Magic byte validation, size limits, file renaming, signed URLs
Data Security	Secrets to env vars, PII scrubbing from logs, encryption at rest, audit logging
Infrastructure	Security headers (CSP, HSTS, X-Frame-Options), strict CORS, HTTPS enforcement
Admin Panel	Auth + role middleware, admin action logging, rate limiting
Payment	Webhook signature verification, server-side price validation, idempotency keys
Supabase	RLS policy audit and fix, Edge Function input validation, cross-tenant prevention

Output format

For each fix applied:

[FIX for VULN-001] Sanitize AI content with DOMPurify
Status: FIXED
What was done: Added sanitizeAIContent() wrapper using DOMPurify with restrictive tag whitelist
Files modified: src/lib/sanitize.ts (new), src/pages/StudySession.tsx, + 3 other pages
Test: Injecting <script>alert(1)</script> into content now renders as plain text
Prevention: All AI content must pass through sanitizeAIContent() before dangerouslySetInnerHTML

Hardening checklist

After fixing all vulnerabilities, the Blue Team generates a final checklist:

- [x] Security headers implemented
- [x] Rate limiting active on auth and API
- [x] CORS restricted to known origins
- [x] Logging and audit trail working
- [x] Sensitive data encrypted at rest and in transit
- [x] Secrets in env vars (none hardcoded)
- [x] Dependencies updated (0 critical/high CVEs)
- [x] RLS policies audited
- [x] Input validation on all endpoints
- [x] File upload secured
- [x] Admin panel protected
- [x] Webhooks verified
- [ ] LGPD compliance (consent, data retention, audit log)

---

Running the Agents

First audit (Red Team)

# In Claude Code
> Run the red-team-scanner agent on this codebase

The Red Team will:

1. Map your stack, endpoints, routes, and middleware
2. Run static analysis for vulnerable patterns
3. Run npm audit for known CVEs in dependencies
4. Review configuration (env vars, CORS, headers)
5. Trace sensitive data flows (PII, tokens, credentials)
6. Generate a numbered report (VULN-001, VULN-002, ...)

Apply fixes (Blue Team)

# In Claude Code, after receiving the Red Team report
> Run the blue-team-defender agent. Here is the Red Team report: [paste report]

The Blue Team will:

1. Triage by severity (CRITICAL and HIGH first)
2. Reverse-engineer each attack vector
3. Implement the fix in code
4. Run tests to verify nothing broke
5. Commit each fix group: fix(security): [VULN-XXX] description
6. Generate a hardening checklist

Re-scan (iterate)

# Run Red Team again to catch regressions or new issues
> Run the red-team-scanner agent again. Previous audit fixed VULN-001 through VULN-012.

Repeat until the Red Team report comes back clean.

Quick commands

Command	What it does
Run red-team-scanner	Full security audit
Run blue-team-defender	Fix vulnerabilities from Red Team report
Run red-team-scanner on src/api/	Scan specific directory
Run red-team-scanner focusing on authentication	Targeted scan

---

Real Vulnerabilities Found

These are real vulnerabilities the agents found in a production SaaS application (React + Supabase + Edge Functions):

Critical

ID	Vulnerability	Impact
VULN-002	Self-referral in affiliate system	Users could generate commissions for themselves
VULN-010	Duplicate commission processing	Same payment could trigger double commission

High

ID	Vulnerability	Impact
VULN-003	XSS in 4 pages via dangerouslySetInnerHTML	Script injection in study content, session theft
VULN-004	API endpoints with zero authentication	Anyone could consume AI quota
VULN-005	Admin routes accessible via race condition	Regular users could access admin panel
VULN-006	Referral field could be overwritten	Affiliate attribution manipulation
VULN-007	Email templates vulnerable to HTML injection	Phishing via manipulated user names
VULN-008	JWT parsed without signature verification	Authentication bypass

Medium

ID	Vulnerability	Impact
VULN-009	No payload limit on bulk import endpoint	DoS via oversized request
VULN-011	CORS wildcard on server-to-server webhook	Cross-origin request abuse
VULN-012	CPF enumeration without authentication	LGPD violation, data harvesting
VULN-013	SQL functions with incorrect volatility	Cache poisoning in query planner

All fixed automatically by the Blue Team agent.

---

Customizing for Your Stack

The agents work out of the box with any stack Claude Code supports. To optimize them for your project:

1. Change the identity line

Replace the project description in both files:

Both agent files start with a generic identity line. Replace it with your project context:

# Default (generic)
You are the **Red Team Scanner** for this project.

# Customized (example)
You are the **Red Team Scanner** for Acme Corp — a B2B fintech SaaS handling payment processing.

2. Add stack-specific sections

If you use Firebase instead of Supabase, replace section 9:

### 9. Firebase-Specific
- Firestore security rules audit
- Firebase Auth configuration
- Cloud Functions input validation
- Storage rules (public vs private)

3. Add compliance requirements

### 10. SOC 2 Compliance
- Access logging for all sensitive operations
- Data encryption at rest and in transit
- Incident response procedures documented
- Vendor security assessment

4. Change the model

For deeper analysis, switch to Opus:

model: opus  # more thorough, higher cost

For faster scans on large codebases:

model: haiku  # faster, cheaper, less thorough

5. Change the language

The agents work in any language Claude supports. Just rewrite the instructions:

# English version
You are the **Red Team Scanner** for this project.
## Identity
- Senior Security Engineer specialized in SaaS, OWASP Top 10, API Security

---

FAQ

Do I need security experience to use these agents? No. The Red Team explains each vulnerability with attack scenarios in plain language. The Blue Team fixes them automatically. You just review the changes.

Will the Red Team modify my code? No. It runs in plan mode (read-only). It can read files, search code, and run npm audit, but cannot write, edit, or execute arbitrary commands.

Will the Blue Team break my app? It runs tests after each fix and commits changes separately. If something breaks, you can revert individual commits. It also runs in acceptEdits mode, so you approve each change.

How much does it cost? The agents use your existing Claude Code subscription or API credits. A full audit of a medium project (~100 files) typically takes 5-15 minutes with Sonnet.

Can I run them in CI/CD? Not directly — Claude Code agents are interactive. But you can schedule periodic audits as part of your security workflow.

Are the agents language-specific? The instructions are in Portuguese but work with any programming language. Claude Code understands the codebase regardless. Feel free to translate the agent files to your preferred language.

What's the difference from tools like Snyk or SonarQube? SAST tools find pattern-based issues (known CVEs, simple code smells). These agents understand your architecture, trace data flows across components, and find logic vulnerabilities (self-referral, race conditions, privilege escalation) that static tools miss.

---

Architecture

.claude/
└── agents/
    ├── red-team-scanner.md      ← ~120 lines, read-only
    └── blue-team-defender.md    ← ~140 lines, edit mode

Workflow:
┌──────────────┐    report    ┌──────────────┐    fixes    ┌──────────┐
│  Red Team    │ ──────────→ │  Blue Team   │ ─────────→ │ Codebase │
│  (scanner)   │             │  (defender)  │            │ (safer)  │
└──────────────┘             └──────────────┘            └──────────┘
       ↑                                                       │
       └───────────────── re-scan ─────────────────────────────┘

---

Credits & Links

• Claude Code — Anthropic's coding agent
• Claude Code Agents — Custom agent documentation
• OWASP Top 10 — Security vulnerability reference

---

If this helped you secure your app, give the repo a ⭐ and share it with other devs shipping fast.