The problem.
To defend well, you need to know how attacks work. I wanted to practise web penetration testing and evaluate an AI-driven pentest agent, without ever leaving the legal frame.
What I built.
01
Frame
- Written rules of engagement: written authorisation required, never production, allow-listed targets, network isolation (no LAN access), AI engine on subscription only.
- Every third-party tool vetted before installation (hidden instructions, data exfiltration, install scripts).
02
Practice
- OWASP Juice Shop as the training target, reset after each test series.
- OWASP ZAP passive, active and browser-driven AJAX scans: SQL injection found and reports archived.
- Next step: compare what the AI pentest agent finds and misses against the target's known flaws.
The architecture.
Test planallow-listed targets
Dedicated serverisolated network
OWASP ZAPpassive · active · AJAX
AI agentunder evaluation
Juice Shoptraining target
Reportflaw · proof · fix
Security & guardrails.
- Written authorisation before any test (French Penal Code art. 323-1).
- Docker network with no LAN access.
- Test target reset, backup before each series.
The results.
- Lab operational with first proven results; AI agent evaluation in progress.
