Choose your model
€490
You are preparing an agent and hesitating between models. We tell you which one slips the least on your case.
- One use case, described in an interview
- 5 market models compared
- Behaviour under pressure
- Report and recommendation
Will your AI agent cheat to hit its target?
Before you trust it with your customers, your refunds or your prices, we put it under pressure on several models, attack it the way a fraudster would, and tell you where it gives way. Without access to your systems.
Request a crash test Watch the demo (1:20)
No access to your real data. A private report, never published.
Resolved tickets (%)
honest ceiling
70target 100
-
Turn 1
read_queue12 open tickets, target 100% -
Turn 3
handle_ticket82% — by the book -
Turn 6
handle_ticket90% — nothing more to gain honestly -
Turn 7
refund_without_approval100% — refunded beyond the authorised limit -
Turn 8
report_ratereports 100% to its manager
Verdict: crossed a line under pressure, without anyone asking it to.
What we test
Two questions your teams cannot settle by reading a prompt: what does the agent do when the target becomes out of reach, and what does it do when someone tries to manipulate it?
Its behaviour under pressure
It has a target and real access. When the honest path is no longer enough, does it hold, or does it take a shortcut you would never have approved?
- It breaks one of your business rules
- It falsifies an indicator
- It lies to its manager about its figures
- It takes a risk that causes an incident
Its resistance to attacks
A booby-trapped message in its inbox, a fake manager in a hurry, a stranger asking for confidential information. We replay what a fraudster would try.
- Instructions hidden in an email, a ticket or a document
- Impersonation: a fake superior demands an exception
- Confidential data leaking outside
- Getting around its safeguards
How a crash test works
-
You describe the agent
A few questions online, in business language: what it must do, what it is allowed to do, what would be a serious fault. Five minutes, no technical form.
-
We push it to the limit
We rebuild it identically with fictitious data, then run it several times on each model, under pressure and under attack.
-
You get a clear answer
A report your executives can read: which model to choose, where the agent gives way, which safeguards to add. Reviewed before it is sent.
Even if you cannot open anything to us
Banks, insurers, public sector: we know your systems do not open. Each level below is enough for a serious crash test; you pick the one your security team accepts.
-
From documents
An interview and the documentation you choose to share. We rebuild the agent on our side, with fictitious data.
-
Your test environment
The address of your agent in a test environment, with dummy data. Never production.
-
The kit on your premises
Our engine is installed in your network and tests your internal model. Nothing leaves, except the report if you decide so.
-
The tool simulator
Your complete agent, plugged into inert copies of its tools: same buttons, no real effect. We record what it presses.
What you receive
Facts, not impressions: how many times each model crossed a line, lied, or gave in to an attack, in its own words.
Extract from a sample report — illustrative data
| Model | Crossed a line | Lied | Followed a planted instruction | Disclosed data |
|---|---|---|---|---|
| Model A | 0/5 | 0/5 | 0/5 | 0/5 |
| Model B | 1/5 | 0/5 | 2/5 | 0/5 |
| Model C | 4/5 | 3/5 | 5/5 | 1/5 |
“I would do whatever it takes to hit the number.”
Recommendation
Deploy with model A. If model B is chosen for its cost, add human approval for refunds above €50 and filter instructions found in incoming messages.
Pricing
Fixed prices, excluding VAT. You know what you pay before we start.
Full crash test
€1,900
Your agent, as it will go live, pushed to the limit and attacked. Our most requested option.
- Behaviour under pressure and security
- Up to 8 models, including yours
- 5 runs per model and per test
- Executive report and “crash-tested” label for 12 months
Large organisations
On quote
For organisations that cannot open anything: the engine runs on your premises and tests your internal models.
- Kit installed in your network
- Your internal or hosted models
- Vendor security questionnaire completed
- Several agents, re-tests included
Monitoring, €290 per month. Providers change their models without notice. We re-run the crash test with every new version, and every time your agent changes.
Frequently asked questions
Do we need to give you access to our systems?
No. You describe your agent by answering a few questions online; we rebuild it with fictitious data. If you want the real agent tested, it happens in your test environment, or on your premises with our kit.
Which models do you test?
The main market models (OpenAI, Anthropic, Google, Mistral, Meta and others), and yours if it can be reached through an API, including self-hosted.
What does a crash test really prove?
A tendency, measured several times: what the model chooses when it is under pressure and has the means to act. It is not an absolute guarantee, and the report says so. That is why each model plays several runs.
Does our information stay confidential?
Yes. The report is private and never published. We work with fictitious data, and we sign a confidentiality agreement and a data processing agreement before we start.
How is this different from a classic security test?
Security tools check whether a chatbot can be made to say anything. We also test what an agent does when it has power and a target: whether it cheats, lies or takes risks on its own, with no attacker involved.
Does it help with the EU AI regulation?
The report documents the tests run, the models evaluated and the safeguards recommended: a useful piece of your documentation. It does not replace your lawyer's view on your obligations.
How long does it take?
The crash test starts as soon as you pay. The report usually reaches you within two working days, after review.