AI Pentesting

Applications with AI agents and AI-enabled applications introduce attack paths that traditional pentests do not cover. We test how attackers can manipulate AI behaviour and connected systems through techniques such as prompt injection, goal hijacking, context poisoning and insecure tool use, and uncover what they can ultimately access, change or execute.

The risks you are facing

AI makes it easier to build, automate and add new capabilities fast, but it also introduces new security risks that developers and security teams need to understand and manage.

AI agents don’t always behave as expected

AI agents make autonomous decisions based on prompts and context, making their behaviour harder to predict or control.

AI agents can be given too many permissions

Agents are given access to APIs, databases and other tools to perform actions autonomously. If these permissions are too broad, an attacker could exploit the agent to perform actions or access systems that the user is not authorised to access.

AI can break data isolation

AI applications process information from users, data sources and connected systems. Without proper isolation, sensitive information from one user, tenant or environment could be exposed to another.

AI turns everyone into a developer

AI makes it possible for more people to build applications, generate code and add functionality without a traditional development background increasing the risk of introducing new vulnerabilities.

How we put your security to the test

We test AI-applications and AI-agents using the OWASP Agentic & GenAI Top 10 framewaroks, covering risks across the full agentic attack surface, from goal hijacking and tool misuse to identity abuse, context poisoning, insecure agent communication and rogue agents.

Define what you want to test

Together with your team, we define the scope, objectives and planning of the pentest. We document these in a test plan, so everyone knows what we will attack, how we will test it and what is off limits.

Prepare for the pentest

Before the pentest begins, we make sure everything is ready for testing. We verify access, test accounts, environments and technical requirements with your team, preventing unnecessary delays and keeping the available test time focused on what matters: finding vulnerabilities.

Identify vulnerabilities

Our pentesters search within the agreed scope for vulnerabilities, AI-specific risks and weaknesses in the application’s business logic. They validate their findings and investigate what an attacker could achieve by exploiting them, while the Lead Pentester remains your dedicated point of contact and keeps you updated throughout the test.

Turn findings into insight

Our pentesters compile the findings, their impact and recommendations into a clear pentest report. An separate pentester reviews the report for quality and accuracy, giving you a reliable basis for remediation and security decisions.

Debrief and define next steps

After the pentest, we walk through the report together and discuss the findings in detail. We provide strategic advice on remediation, giving your team a clear understanding of the findings and how to address them.

What we pentest

From a single AI chatbot to complex applications with AI agents: the scope determines which models, agents, integrations and underlying systems we put to the test.

AI agents and agentworkflows

We try to steer agents away from their intended goals. Using techniques such as prompt injection, goal hijacking and context manipulation, we test how much control an attacker can gain over their decisions and actions.

Tools and platforms for AI agents

The environments where users build and configure agents can also be targeted. We test the boundaries between users, permissions, tools and instructions, looking for ways to break through them.

AI chatbots

We push chatbots beyond their intended boundaries. Can instructions be bypassed, sensitive information be extracted or functions be triggered that were never intended for the user?

AI agent architecture review

Here, we focus on the design behind the agent. We assess data sources, permissions, tools, models and underlying systems to uncover vulnerabilities early in the design process.

AI integrations in Mendix

With the introduction of Maia and AI agents, we combine our Mendix and AI security expertise to test how attackers can exploit new AI functionality in Mendix applications.

System Prompt Extraction

We try to expose hidden system instructions. Using carefully crafted inputs, indirect prompt injection and other techniques, we test whether an attacker can retrieve or reconstruct confidential instructions.

Your primary point of contact

“Hacking AI is never the same twice: what seems secure today could introduce a new vulnerability tomorrow, so we stay on top of developments and continuously sharpen our attack techniques and advice.”

Nick Bruinsma

Director Innovation

FAQ

How long does a pentest take?

That depends on the size, complexity and scope of the environment being tested. During scoping, we determine what needs to be tested and agree on a realistic planning before the test starts. Like a regular pentest, an engagement can range from 2 to 50 days.

What is the difference between a standard pentest and an AI pentest?

A standard pentest assesses the security of applications, APIs and underlying systems. An AI pentest does the same, but also tests how models, agents, prompts, knowledge sources and connected tools can be manipulated. We look beyond unexpected AI behaviour to determine whether it could lead to unauthorised access, data leakage or unintended actions.

What do you need to get started?

We start with a clear scope, objectives and a technical point of contact. Depending on the test, we may also need test accounts, access to specific environments or other information required to realistically test the agreed scope.

Take your pentest further

Red Teaming

Discover how far an attacker can get in a realistic attack scenario with a specific objective. This takes Red Teaming beyond the defined scope of an AI pentest.

Security Awareness Training

Make AI risks tangible for employees with practical training based on relevant AI pentest findings and attack techniques.

Security Health Check

Get a clear view of how your organisation secures AI. We assess areas including access control, shadow IT, third-party AI services, agent permissions, data processing and incident response.

“The S-Unit combines in-depth expertise with a pragmatic approach. They think along critically, communicate clearly and translate their findings into concrete improvements we can act on straight away. It’s a pleasant way of working that helps us strengthen our information security further. And just as important: thanks to them, I sleep a lot better”

Firas Daoud

TAF – CTO

“The S-Unit understood from the start what we wanted to achieve, even though the path to get there wasn’t entirely clear yet. They took us along in their approach, thought ahead and switched gears quickly when needed. We could trust that it would work out, without having to keep steering. It’s a really pleasant way of working together.”

Martijn Korevaar

Gemeente Rotterdam – IT Risk & Compliance Manager

“We didn’t want to be told that we’re vulnerable. We wanted to see how far a real attacker would get. Instead of a standard report, we got insight into exactly the risks that matter most to us.”

Anonymous

Manager Tech Security

“We received professional help with a pentest of a broad application landscape. Despite the many different functionalities in our landscape, the test was fast and thorough.”

Daan van der Ham

Lead techniek - PO-VO Raad

Ready to put AI to the test?

Whether you already have a scope in mind or are still figuring out what to test, we’ll help define the right offensive approach.