All servicesAI red team
Find out what your AI gives away before a stranger finds out for you.
An integrity audit of the AIs you already use: customer chatbots, internal assistants, vendor tools. We test them, then hand you a written report that says where they hold and where they give way.
On quote.
Your AI, as deployed
We test what you already run, in its real setup, not a demo.
Three questions
What it leaks, what it can be made to say or do, and what it does with your data.
A written report
Every finding comes with the test that shows it, its severity and the recommended fix.
Rules set first
Scope and limits are agreed in writing before the first test.
Your chatbot talks to your customers. Do you know everything it can say?
You connected an AI to your website, your email or your documents, and it is useful. But nobody on your side has tried to push it out of its role. The vendor's data sheet describes the product, not the way it is wired into your organization.
| Trusting the vendor | Poking at it yourself | With Cerbère | |
|---|---|---|---|
| What gets tested | The product in general, by the company selling it | A few trick questions, at random | Your actual setup, against a written test plan |
| Data leaks | Described in a policy, not checked on your setup | Hard to spot without a method | Targeted tests: what the AI reveals about your documents, your customers, its own instructions |
| Booby-trapped text | Outside the product sheet | Requires knowing the techniques | Direct injection tests, and indirect ones through content the AI reads |
| What the AI can do | Whatever permissions were granted at setup | Permissions rarely reviewed | A review of what the AI can do with the access it was given |
| What you end up with | A general assurance | A hunch | A written report: where it holds, where it gives way, what to fix first |
A test does not prove an AI is safe. It shows where it gives way today, while there is still time to fix it.
Prompt injection is entry number 1 (LLM01) in OWASP's 2025 list of risks for large language model applications.
6 cyber security agencies, including the Canadian Centre for Cyber Security, state: 'Privileges assigned to agents directly determine the level of risk they can introduce.'
What we test
Four angles, one report
What it leaks
Internal instructions, excerpts of your documents, details about other people: we look for what the AI reveals when pressed.
What it can be made to say
We try to push it out of its role: statements that commit you, wrong answers delivered with confidence, ways around its limits.
What it can be made to do
If it can send, change or delete, we check whether text planted in a web page or a file can trigger those actions.
What it does with your data
We read the settings and the vendor's terms: what is kept, and what may be used to train a model.
Who it is for
A small business with a customer chatbot
An assistant answers in your name on your site or in your messages. What it says commits you, and it may reach more documents than you think.
A practice, clinic or non-profit with an internal assistant
Your team queries an AI connected to case files. You want to know whether someone can read what they should not see.
This is for you if
- An AI answers your customers or your team in your name.
- An AI has access to your email, your documents or your customer records.
- You hold personal information and must be able to say what happens to it.
- You want written findings before rolling an AI tool out to the whole team.
Probably not for you if
- You use no AI connected to your data or your customers.
- You want a certificate declaring your AI safe. No test can support that statement.
- You want to fence in an agent that acts on your systems, not test one: see Trustworthy AI agents.
What is included
What is included
Scope and rules of engagement
Which AIs, which accounts, which limits. Set in writing before the first test.
Access inventory
What each AI is connected to: documents, email, tools, possible actions.
Leak tests
Attempts to extract internal instructions, other users' data, or content that should not come out.
Injection tests
Direct injection in the conversation, and indirect injection through content the AI reads.
Settings and terms review
What the settings and the vendor's terms provide for your data: retention, training, sharing.
Written report
Every finding with the test that reproduces it, its severity and the recommended fix, in priority order.
Questions in writing
After delivery, you send your questions in writing and we answer in writing.
How it goes
How it goes
-
Step 01
You write
Tell us which AIs you use and what they are connected to. Written reply within 24 business hours.
-
Step 02
We scope
Scope, rules of engagement and quote, in writing. No test before you agree.
-
Step 03
We test
Tests follow the agreed plan, on the agreed accounts, and touch nothing else.
-
Step 04
We report
The written report says where your AIs hold, where they give way, and what to fix first.
Pricing
Price
-
AI red team
On quote.
- Written scope and rules of engagement
- Leak, injection and hijacking tests
- Review of settings and vendor terms
- Written report, fixes in priority order
The quote depends on how many AIs you run and what they can reach. It covers the tests and the report, not applying the fixes.
What we do. What we don't.
We do
- Put everything in writing: scope, rules of engagement, report.
- Test only what you authorized in writing.
- Leave your accounts and access in your hands.
- Give you, for every finding, the test that lets you reproduce it.
We don't
- Declare an AI safe.
- Test a system without consent from whoever must give it.
- Inflate a minor finding to sell the next job.
- Tie you into a contract you cannot leave.
Frequently asked questions
What is an AI red team?
A team that plays the adversary against your AIs: it tries to make them reveal, say or do what they should not, then reports in writing what worked and what held.
My vendor says its AI is secure. Why test?
The vendor is talking about its product. The test is about your setup: the documents you connected, your instructions, the permissions you granted. Those sit on your side, and you are the one who can fix them.
What is prompt injection?
Text that makes the AI behave in ways nobody intended. It can be direct, typed into the conversation, or indirect, hidden in content the AI reads, such as a web page or a file. OWASP lists it as number 1 in its 2025 list.
Can the tests harm my service?
Limits are set in writing before we start: which accounts, which actions. Where possible, we test on a trial account instead of real customer data.
Can you test a tool that belongs to a vendor?
We test your use of the tool, with your accounts, within the vendor's terms of use. If a test needs the vendor's consent, we tell you and wait for it.
What is in the report?
For each finding: what was tried, what the AI answered or did, the severity, and the recommended fix. The report also says what held.
What if you find a leak of personal information?
We tell you in writing right away, without waiting for the final report. In Quebec, a confidentiality incident that presents a risk of serious injury must be reported to the Commission d'accès à l'information and to the people concerned. Assessing that duty is yours, with the person in charge of personal information in your organization.
Once the fixes are in, will my AI be safe?
No. OWASP itself writes that it is unclear whether fool-proof methods of prevention exist for prompt injection. Fixes reduce the risk. They do not remove it.
What does it cost?
On quote. The price depends on how many AIs are tested and what they are connected to. You get the quote in writing before any test.
Tell us which AIs you use.
Name the tools and what they can reach. Written reply within 24 business hours: what is worth testing, and what is not.
Written reply within 24 business hours. No video calls, no booking calendar.