My AI employee
Falk tests my application like an attacker and tells me what to fix first.
Falk has been running for me since August 31, 2026, and has completed one authorized engagement so far. There's no ready-made package for him yet in the community, because every engagement needs its own scope and its own approval.
Falk tests an application in stages: from the outside with no access at all (recon), then as an unauthenticated attacker from the outside (blackbox), and, where I approve it, with access or repo insight (greybox, whitebox). For every finding he delivers evidence, a CVSS severity rating and a concrete fix. He only tests systems that belong to me or for which written approval exists.
A real workday
One engagement is completed so far, on an application of my own, external and with no prior access. What's here is the method, not the findings themselves: details about the target or the outcome stay in the internal engagement folder for good reason.
4
Testing stages in the model
2 of 4
Completed so far
max. 20 requests/min
Scan pace
Before a single request went to the target system, my written approval was in place. For Falk, that's a fixed gate, not a formality. The process followed five phases: first, reconnaissance, where he derives the technology in use strictly from publicly served files instead of guessing, then an attack surface model along the login and session flow, then manual tests for typical vulnerability classes like open redirects or information leaks in error messages, a tooling check of TLS configuration and an automated scan, and finally the assessment.
Every finding comes with a CVSS severity rating and a concrete, prioritized fix, not just a list of issues. The automated scan ran deliberately throttled: four parallel requests, capped at 20 per minute, no load or stress testing, no trying of credentials. Rate-limit etiquette toward the target system isn't an afterthought for Falk, it's a fixed part of the scope.
The tools he works with
Falk works with a staged toolkit, depending on which stage is approved for the engagement:
- Passive recon tools that only read publicly served files and headers, no request with a side effect.
- Manual review of the login and session flow directly in the frontend code, no automated assumptions.
- A targeted, rate-limited automated scanner for TLS configuration and known vulnerability classes.
- A written scope and an engagement folder per job, logging every approval and every step.
What he can't do, where I approve
- He only tests systems that belong to me or for which written approval exists. No test without scope approval in the job order.
- Active scans, fuzzing or exploit attempts against a target system are gated, as is spending money, for example on commercial tools or cloud scan quotas.
- Never: third-party systems without a written scope, destructive actions like outages or data deletion, exfiltration of real personal data, quoting found credentials in plain text.
- Every engagement has fixed abort criteria: instability, unexpected data access, a third-party target, or rate-limit warnings stop the test immediately.
- A findings report goes through a review pass by a separate reviewing role before I see it.
The path: build it yourself, there's no ready-made template
Standard: build it yourself
Entry point is Learning Claude, plus the articles on hiring and creating an AI employee. What's special about Falk: he thinks from an attacker's perspective instead of just ticking off checkpoints, he works in stages instead of firing tools at random, and every finding comes with a CVSS rating and a concrete fix, not just a warning.
Honestly: not a package
Falk only runs in my own business. There's no ready-made template yet, because a security test always needs its own scope and its own approval for the given target system, that can't be generalized into a template.
Frequently asked questions about Falk
How do I check if my website is secure?
A single check rarely covers it. A staged approach makes sense: first from the outside with no access, then as an unauthenticated attacker, and where it makes sense, with real access or repo insight. Free online checkers give a first impression, but they don't replace a targeted, manual check with evidence and a severity rating.
Can AI help with penetration testing?
Yes, in parts. AI can speed up recon, scan code for typical vulnerability patterns, and structure reports. For me, Falk works in stages and documents every step before acting. What he can't do: act on his own against third-party systems. Every test needs a written scope and my approval.
How do I test my own website for security flaws?
The first step is reconnaissance based strictly on publicly served files, with no assumptions. Next comes an attack surface model along the login and session flow, then manual tests for typical vulnerability classes and a rate-limited automated scan. At the end sits an assessment with a fix proposal for every finding, not just a list of errors.
What does a penetration test cost?
That depends heavily on the scope: the size of the application, which stages are approved, and whether it stays at an external blackbox check or also includes access or repo insight. For me this runs within my own business, so I can't honestly give a flat number without knowing the actual scope.
What's the difference between an automated scan and a real red team test?
An automated scan looks for known patterns and reports hits without understanding the context. A red team test like Falk's combines that with manual review, an attack surface model, and a severity rating with a fix proposal. The scan is one tool inside it, not the whole test.
Finding security flaws before an attacker does
There's no template for Falk's exact setup, but the learning path he grew out of is in the community.
Join the community