PT/04 · Penetration testing
Find out what your AI will do for an attacker.
AI assistants can read your data and take actions on your behalf. We test what they could be persuaded to reveal or do, before someone else finds out.
01Overview
What we look at.
We start by understanding how your AI system is set up: what it can see, who uses it, and what it's allowed to do. That's where the real risk lives, so that's where the testing goes.
Then we try to turn it against you. AI behaves differently from run to run, so every finding records how reliably it happened, not just that it happened once.
In scope
- Whether the AI can be talked out of its rules
- Leaks of confidential data, documents or instructions
- Actions the AI could be tricked into taking
- The application and access controls around it
02Why test
What you get out of it.
Find AI-specific risks
These problems don't show up in a traditional penetration test. They need testers who know how AI systems fail.
Limit the damage
The best fixes are often about access and approvals, not the model itself. We recommend the ones that fit your setup.
Support your AI governance
Customers, boards and insurers are starting to ask how AI risk is managed. A test report is solid evidence.
Launch with confidence
Before launch is the cheapest time to learn what your AI will do when someone pushes it.
Ground rules
- Usage and cost limits agreed up front and watched during testing
- We test your configuration and data, not the vendor's model
- If anything looks unstable, we stop and call you
03Included
Every engagement includes.
See a sample reportA clear, written report
A summary for leadership, then each finding with evidence, a severity rating and a fix your team can act on.
Critical findings, same day
If we find something critical, we call you that day instead of waiting for the report.
Retest included
Once fixes are in, we retest at no extra cost and update the report with what's closed.
Attestation letter
A one-page summary of scope, dates and outcome for customers, auditors and insurers.
Readout call
We walk your team through the results and agree what to fix first.
Your data, handled carefully
Test evidence stays encrypted and is deleted after the retest unless you ask us to keep it.
04FAQ
Common questions.
How long does an AI system test take?
Usually one to two weeks of testing, depending on how many AI features and data sources are involved. The report follows within a week.
What does an AI system test cost?
Every test gets a fixed quote after a free scoping call, based on what's actually in scope. No hourly billing and no surprises on the invoice.
How is this different from a normal penetration test?
A traditional test looks for flaws in software and infrastructure. AI testing adds the model's behaviour: whether it can be persuaded to ignore its rules, share what it knows, or misuse what it has access to.
Do you test the regular application too?
Yes, wherever the AI touches it. A full web application test can be added to the same engagement.
Can you test Microsoft 365 Copilot or ChatGPT Enterprise?
Yes, as you've configured them: what they can reach in your data and what they surface to whom. We test your setup, not the vendor's model.
Other testing
PT/01
Network & Microsoft 365 testing
What an attacker can reach from the internet, how far they could get once inside, and how well your accounts hold up.
PT/02
Web application testing
Your customer portals, web apps and integrations, tested by people trying to make them do what they shouldn't.
PT/03
Social engineering
Realistic phishing, phone and text campaigns that show how your team responds, followed by training aimed at what worked.
Next step
Scope your test.
You'll speak with the person who does the testing. We reply within one business day.