AI red teaming: find the gaps in your AI before others do

With an AI red team test, you find the gaps in your chatbot or AI agent before an attacker or a customer does. We attack your system the way an outsider would, and you get every finding with a priority and a concrete fix. Afterward, you go live with confidence and have evidence for NIS2, the EU AI Act and for customers who ask.

What you get from it

Your sales team can answer the customer's security questionnaire.
Large customers ask before the contract how you secure your AI. You hand over the report instead of stalling.
Your chatbot doesn't reveal internal prices.
Before a resourceful visitor gets it to give away price lists, discount rules or internal instructions, you know about it.
Your agent sends nothing that nobody wanted.
A prepared email doesn't get it to forward invoices or change records.
Customer data stays where it belongs.
Nobody gets to see other customers' data or personnel files through the chat window.
You have something in hand for the audit.
Every finding is mapped to the matching obligation under NIS2, GDPR or the EU AI Act.

What your next website could look like

See all work →

What is AI red teaming?

AI red teaming is a targeted attack on an AI system to find its weaknesses before they do damage. A classic penetration test asks whether someone can get into the system. An AI penetration test also asks what the system does when someone manipulates it with language. An attacker doesn't need server access for that. Often a single sentence in the chat window, in an email or on a website your agent reads is enough.

In the intro call we set your goal together.

You learn which first step is worth it for you.

Which risks we uncover for you

Internal prices
A customer gets the chatbot to reveal internal prices, margins or discount rules.
Hidden instructions
An email with hidden instructions gets the agent to forward invoices to an outside address.
Damage to your brand
A visitor talks the AI into statements that harm your brand or are legally sensitive.
Access that is too broad
An employee asks the internal assistant and gets to see salary lists or contracts not meant for them.
Leaked instructions
Someone extracts your bot's internal instructions and copies your offer.
Too many permissions
An agent is allowed to do more than its task requires, and a single error triggers a payment or deletion.

What we align with

We align the attacks with the well-known risk lists, above all the OWASP LLM Top 10. Our article on agentic attacks describes how attacks on agents work.

Why Aqtelo

Every finding comes with a priority and a fix.
You don't get a list of warnings, you know what to do first and how.
Findings are mapped to your obligations.
We map every finding to NIS2, GDPR or the EU AI Act. That saves you the translation for management and auditors.
We build AI systems ourselves.
Our Intelligence area builds agents and chatbots. We know where they break, because we have to secure them ourselves.
We also test systems from other providers.
It makes no difference whether your team, an agency or a software vendor built the system.
One fixed contact.
You speak with one person, in German, English or Portuguese. Experienced specialists stand behind them.
Your data stays in the EU.
It trains no third-party model. Which data we may see, we set in advance.

How we work together

  1. Set the scopeWe clarify which system we test, which data and actions it uses and what is excluded. We put that in writing.
  2. AttackWe attack the system and record every successful attack with input and result.
  3. Hand over the reportYou get every finding with priority, example, fix and a link to NIS2, GDPR or the EU AI Act.
  4. RetestOnce the fixes are in place, we test again. That way you see that the gap is closed.

In the intro call we set your goal together.

You learn which first step is worth it for you.

Classic penetration tests too

We also offer classic penetration tests for websites and interfaces. Security always tests systems delivered by Aqtelo before launch, whether AI agent or AI chatbot.

Reference websites

All work

Intro call

Let's find out together where your growth gets stuck

  • You talk to the team that also does the work. No sales loop.
  • You learn which first step is worth it for you.
  • If your question fits another specialist better, we'll tell you.
Renata WisniewskaManaging Director, AqteloI read every request myself and reply within one business day.office@aqtelo.com

Start a project

What can we help with?

You can select more than one

What you probably want to know

It depends on the system, its data and actions and the scope of the tests. After the first call, we define the scope. Then you get a fixed quote.

It depends on the scope. We set the schedule together with the scope, including a date for the report. We plan the retest once your fixes are done.

Access to the system, ideally in a test environment, and a contact person for questions. Plus a short description of what the system may do and which data it uses.

Where possible, we test in a test environment with test data. If testing has to happen in the live system, we agree time windows and limits with you beforehand.

A report with every finding, its priority, an example and the fix. Every finding is mapped to NIS2, GDPR or the EU AI Act. After the retest, you see which gaps are closed.

No. The report helps you demonstrate obligations and talk to auditors. Which obligations apply to your company is for your law firm to clarify.

Keep reading