AI systems and security review.
I look for what breaks an AI system before a client finds it.
I review an AI agent, flow or website that is already live, or about to go live. Two independent reviewers, every finding with the proof that reproduces it, its severity and its fix. What gets fixed is reviewed again before it is closed. The scope is agreed in writing and so is the price: whatever turns up changes the report, not the invoice.
- Title block
- Takes
- Two weeks · one system
- You get
- A report with findings by severity, each with its proof and its fix
- Afterwards
- Re-test included, up to 30 days later
- Price
- €2,900 + VAT, fixed
// Every finding is reproduced on your system. If it cannot be reproduced, it does not go in the report. No title or certification I do not hold: each review is proven by its report.
Where an AI system breaks.
- 01 A · What comes in Messages, documents, forms WhatsApp · Instagram · web
- 05 B · The model The agent and its LLM calls instructions · retries · spend
- 020304 C · Data and tools Database, calendar, CRM, payments who sees what · what it can do
- 06 D · What goes out Replies, notices, confirmations what the client sees
- 01 Prompt injection Messages and documents with hidden instructions: does the agent obey them?
- 02 Data leaking between clients Can one client see or ask for another client’s data, through the chat or the API?
- 03 Permissions What can the system do without anyone asking: read, delete, charge?
- 04 Secrets in the code Keys and tokens in the repository, the browser or the logs.
- 05 Uncapped LLM costs Loops, retries and abuse: is there a cap before the bill arrives?
- 06 Forms that confirm without sending Does the page say “thanks” even when the submission failed?
Five steps, always the same.
- 01
Scope
Which system, which access and how far. In writing, before touching anything. What is out of scope stays untouched.
- 02
Review
Two independent reviewers; one of them a different model in a clean session.
- 03
Executed proof
Every finding, with the proof that reproduces it and its saved output.
- 04
Report
Findings by severity, each with its fix. What costs you most comes first.
- 05
Re-test
What was fixed is reviewed again with the same proof. Included, up to 30 days later.
What you get, in full.
Summary, scope, findings by severity and what was fixed. No client name.
| ID | Type | Severity | Status |
|---|---|---|---|
| F1 | Form | High | Fixed |
| F2 | Form | High | Fixed |
| F3 | High | Fixed | |
| F4 | Token | Medium | Fixed |
| F5 | Body | Medium | Fixed |
| F6 | Medium | Fixed | |
| F7 | Deploy | Medium | Fixed |
| F8 | Storage | Medium | Fixed |
| F9 | Deploy | Medium | Fixed |
| r7 | Quick check | High | Fixed |
| r9 | Quick check | Medium | Fixed |
Review library.
| Date | System | Type | Review | Report |
|---|---|---|---|---|
| LinkedIn outreach with human approval Own system | AgentsHuman approval | 7 rounds | PDF · Pending | |
| Contact form and quick check on this website Own system | WebForms | 9 rounds · fixed; 2 accepted residuals |
What I don’t do.
- Review without written permission Scope and access are agreed first. What is out of scope stays untouched.
- Break anything for real Deleting, charging or sending is simulated. Nothing runs on your live system.
- Publish a report with your name Only anonymised and with the client's permission. In the library: sector and system.
- Close a finding without re-testing What was fixed is reviewed again with the same proof that found it.
- Display badges I do not hold No certification on the shop window. Each review is proven by its report and its date.
Before we start.
Why is a classic pentest not enough?
A pentest looks at the website and the server. This also looks at what the model does with what it is told, what it can touch unasked, and how much it spends.
What if you find nothing serious?
Then you have proof that it holds, in writing and dated. That is the product too.
What do you need from me?
Read access to the code and a test environment, and an hour with you to agree the scope. Nothing else until the report.
Do you have an AI system in production?
Tell me what it does and what it touches: data, payments, clients. I will tell you what I would review first and whether a review is worth it for you.