AI agents generate code fast but not always securely. LLMs can produce code with:

• SQL injection -- Unsanitized inputs in database queries
• XSS -- Unescaped output in HTML templates
• Command injection -- Unvalidated input passed to shell commands
• Hardcoded secrets -- API keys or passwords embedded in code
• Insecure dependencies -- Outdated or vulnerable packages
• Broken authentication -- Weak session handling, missing auth checks
• Path traversal -- Unsanitized file paths
• SSRF -- Unvalidated URLs in server-side requests

These are not hypothetical. They happen regularly in AI-generated code.

# Install pre-commit hook (per OCPA)
ln -sf ../../scripts/pre-commit .git/hooks/pre-commit

In CI:

- uses: gitleaks/gitleaks-action@v2

Never trust that an agent won't commit secrets. This gate is non-negotiable.

- uses: aquasecurity/trivy-action@master
  with:
    scan-type: 'fs'
    security-checks: 'vuln'

Or use GitHub's built-in Dependabot for automatic dependency update PRs.

A second AI agent reviews diffs specifically for security (see AI Review via n8n and n8n Workflow 2):

Review this diff for OWASP Top 10 vulnerabilities.
Check for: SQL injection, XSS, command injection, hardcoded secrets,
insecure deserialization, broken access control, SSRF.
Flag any finding with severity (critical/high/medium/low).

This is not a replacement for human security review. It catches obvious issues before a human looks.

Always require human review for changes touching:

• Authentication / authorization logic
• Payment / billing code
• Cryptography usage
• Infrastructure / deployment configuration
• API endpoint access control
• PII / sensitive data handling

Mark these paths in your CLAUDE.md so the agent flags them:

## Security-Critical Paths (always require human review)
- `app/auth/` -- Authentication and authorization
- `app/payments/` -- Payment processing
- `k8s/` -- Infrastructure configuration
- `scripts/deploy*.sh` -- Deployment scripts

• Use Content Security Policy headers
• Enable rate limiting on all public endpoints
• Run containers as non-root users (configure in Dockerfile)
• Use read-only filesystem mounts where possible
• Set network policies to restrict container communication

Agent generates code
  > Gitleaks (secrets)     -- FAIL > block, agent fixes
  > SAST scan              -- FAIL > block, agent fixes
  > Dependency scan        -- FAIL > block, agent fixes
  > AI security review     -- FLAGS > route to human
  > Touches critical path? -- YES > route to human
  > All clear              -- PASS > continue to deploy

Never fully trust AI-generated code for security. Use tooling to catch the obvious, use AI to catch the subtle, and use humans to validate the critical. The cost of a security breach vastly exceeds the cost of a human review.